Transfer of Learning

With the rapid development of textual information on the web, sentiment analysis is changing to an essential analytic tool rather than an academic endeavor and numerous studies have been carried out in recent years to address this issue. By the emergence of deep learning, deep neural networks have attracted a lot of attention and become mainstream in this field. Despite the remarkable success of deep learning models for sentiment analysis of text, they are in the early steps of development and their potential is yet to be fully explored. Convolutional neural network is one of the deep learning methods that has been surpassed for sentiment analysis but is confronted with some limitations. Firstly, convolutional neural network requires a large number of training data. Secondly, it assumes that all words in a sentence have an equal contribution to the polarity of a sentence. To fill these lacunas, a convolutional neural network equipped with the attention mechanism is proposed in this ...

Sentiment analysis is known as one of the most crucial tasks in the field of natural language processing and Convolutional Neural Network (CNN) is one of those prominent models that is commonly used for this aim. Although convolutional neural networks have obtained remarkable results in recent years, they are still confronted with some limitations. Firstly, they consider that all words in a sentence have equal contributions in the sentence meaning representation and are not able to extract informative words. Secondly, they require a large number of training data to obtain considerable results while they have many parameters that must be accurately adjusted. To this end, a convolutional neural network integrated with a hierarchical attention layer is proposed which is able to extract informative words and assign them higher weight. Moreover, the effect of transfer learning that transfers knowledge learned in the source domain to the target domain with the aim of improving the perform...

A grand challenge in brain decoding is to develop algorithms that generalize across multiple subjects and tasks. Here, we developed a new computational framework to minimize negative transfer for domain-adaptive brain decoding by reframing source selection as a mixture model parameter estimation problem, allowing each source subject to contribute through a continuous mixture weight rather than being outright included or excluded. To compute these weights, we developed a novel convex optimization algorithm based on the Generalized Method of Moments. By using model performance metrics as the generalized moment functions in our GMM optimization, our algorithm also provides theoretical guarantees that the mixture weights are an optimal approximation of the importance weights that underlie domain adaptation theory. When tested on a large-scale brain decoding dataset (n=105 subjects), our new mixture model weighting framework achieved state-of-the-art performance—increasing accuracy up to 2.5% over baseline fine-tuning, double the performance gain compared to previous research in supervised source selection. Notably, these improvements were achieved using significantly less training data (i.e., 62% smaller effective sample sizes), suggesting that our performance gains stem from reduced negative transfer. Collectively, this research advances toward a more principled and generalizable brain decoding framework, laying the mathematical foundation for scalable brain-computer interfaces and other applications in computational neuroscience.

The coronavirus outbreak has caused a devastating effect on people all around the world and has infected millions. The exponential escalation of the spread of the disease makes it emergent for appropriate screening methods to detect the disease and take steps in mitigating it. The conventional testing technique involves the use of Reverse-Transcriptase Polymerase Chain Reaction (RT-PCR). Due to limited sensitivity it is more prone to providing high false negative rates. Also due to a high turnaround time (6-9 hours) and a high cost, an alternative approach for screening is called for. Chest radiographs are the most frequently used imaging procedures in radiology. They are cheaper compared to CT scans and are more readily available and accessible to the public. Application of advanced artificial intelligence (AI) techniques coupled with radiological imaging can be helpful for the accurate detection of this disease. In this paper 8 different architectures of CNN (EfficientNet, ResNet, DenseNet, VGG, Mobilenet) are compared regarding classification performance on the COVID-19 dataset. The proposed models are developed to provide accurate diagnostics for multiclass classification (Covid vs No Findings vs Pneumonia) and the best performing models are used to perform binary classification (Covid vs No Findings). Areas under the receiver operating characteristics curves (AUROC) between 0.9622 and 0.9987 could be achieved for multiclass classification while accuracy scores ranged from 90.32% to 94.93% achieving next to state-of-the-art results. A dataset was created as a mix of publicly available X-ray images from patients with confirmed COVID-19 disease, common bacterial pneumonia and healthy individuals.

Banana crops play a pivotal role in securing global food supplies and supporting economic stability. However, they are confronted with significant challenges stemming from a variety of diseases that not only diminish yields but also compromise the quality of the fruit. Artificial intelligence, especially deep learning, assumes a pivotal role in tackling this challenge by leveraging advanced algorithms and data analysis techniques to enhance disease detection and diagnosis in banana crops, thus contributing significantly to their protection and preservation. To address this challenge, we present the "Banana Leaf Spot Diseases (BananaLSD) Dataset" comprising images of major banana leaf spot diseases and healthy leaves, meticulously labelled by plant pathologists. Using deep learning models, including DenseNet-201, EfficientNet-b0, and VGG16, we achieved remarkable disease classification accuracy rates. DenseNet-201 achieved an impressive 98.12% accuracy. The study analyses performance metrics and visualization by grad-cam technique. These results underscore the potential of deep learning for precise banana leaf disease diagnosis, offering significant implications for crop preservation, economic stability, and global food security.

Transfer learning entails taking an artificial neural network (ANN) that is trained on a source dataset and adapting it to a new target dataset. While this has been shown to be quite powerful, its use has generally been restricted by architectural constraints. Previously, in order to reuse and adapt an ANN's internal weights and structure, the underlying topology of the ANN being transferred across tasks must remain mostly the same while a new output layer is attached, discarding the old output layer's weights. This work introduces network-aware adaptive structure transfer learning (N-ASTL), an advancement over prior efforts to remove this restriction. N-ASTL utilizes statistical information related to the source network's topology and weight distribution in order to inform how new input and output neurons are to be integrated into the existing structure. Results show improvements over prior state-of-the-art, including the ability to transfer in challenging real-world datasets not previously possible and improved generalization over RNNs trained without transfer.

Prostate Cancer (PCa) is the second most common cancer in men and affects more than a million people each year. Grading prostate cancer is based on the Gleason grading system, a subjective and labor-intensive method for evaluating prostate tissue samples. The variability in diagnostic approaches underscores the urgent need for more reliable methods. By integrating deep learning technologies and developing automated systems, diagnostic precision can be improved, and human error minimized. The present work introduces a three-stage framework-based innovative deep-learning system for assessing PCa severity using the PANDA challenge dataset. After a meticulous selection process, 2699 usable cases were narrowed down from the initial 5160 cases after extensive data cleaning. There are three stages in the proposed framework: classification of PCa grades using deep neural networks (DNNs), segmentation of PCa grades, and computation of International Society for Urological Pathology (ISUP) grades using machine learning classifiers. Four classes of patches were classified and segmented (benign, Gleason 3, Gleason 4, and Gleason 5). Patch sampling at different sizes (500 × 500 and 1000 × 1000 pixels) was used to optimize the classification and segmentation processes. The segmentation performance of the proposed network is enhanced by a Self-organized operational neural network (Self-ONN) based DeepLabV3 architecture. Based on these predictions, the distribution percentages of each cancer grade within the whole slide images (WSI) were calculated. These features were then concatenated into machine learning classifiers to predict the final ISUP PCa grade. EfficientNet_b0 achieved the highest F1-score of 83.83% for classification, while DeepLabV3 + architecture based on self-ONN and EfficientNet encoder achieved the highest Dice Similarity Coefficient (DSC) score of 84.9% for segmentation. Using the RandomForest (RF) classifier, the proposed framework achieved a quadratic weighted kappa (QWK) score of 0.9215. Deep learning frameworks are being developed to grade PCa automatically and have shown promising results. In addition, it provides a prospective approach to a prognostic tool that can produce clinically significant results efficiently and reliably. Further investigations are needed to evaluate the framework's adaptability and effectiveness across various clinical scenarios.

Te increasing prevalence of colon and lung cancer presents a considerable challenge to healthcare systems worldwide, emphasizing the critical necessity for early and accurate diagnosis to enhance patient outcomes. Te precision of diagnosis heavily relies on the expertise of histopathologists, constituting a demanding task. Te health and well-being of patients are jeopardized in the absence of adequately trained histopathologists, potentially leading to misdiagnoses, unnecessary treatments, and tests, resulting in the inefcient utilization of healthcare resources. However, with substantial technological advancements, deep learning (DL) has emerged as a potent tool in clinical settings, particularly in the realm of medical imaging. Tis study leveraged the LC25000 dataset, encompassing 25,000 images of lung and colon tissue, introducing an innovative approach by employing a self-organized operational neural network (Self-ONN) to accurately detect lung and colon cancer in histopathology images. Subsequently, our novel model underwent comparison with fve pretrained convolutional neural network (CNN) models: MobileNetV2-SelfMLP, Resnet18-SelfMLP, DenseNet201-SelfMLP, InceptionV3-SelfMLP, and MobileViTv2_200-SelfMLP, where each multilayer perceptron (MLP) was replaced with Self-MLP. Te models' performance was meticulously assessed using key metrics such as precision, recall, F1 score, accuracy, and area under the receiver operating characteristic (ROC) curve. Te proposed model demonstrated exceptional overall accuracy, precision, sensitivity, F1 score, and specifcity, achieving 99.74%, 99.74%, 99.74%, 99.74%, and 99.94%, respectively. Tis underscores the potential of artifcial intelligence (AI) to signifcantly enhance diagnostic precision within clinical settings, portraying a promising avenue for improving patient care and outcomes. Te synopsis of the literature provides a thorough examination of several DL and digital image processing methods used in the identifcation of cancer, with a primary emphasis on lung and colon cancer. Te experiments use the LC25000 dataset, which consists of 25,000 photos, for the purposes of training and testing. Various techniques, such as CNNs, transfer learning, ensemble models, and lightweight DL architectures, have been used to accomplish accurate categorization of cancer tissue. Various investigations regularly show exceptional performance, with accuracy rates ranging from 96.19% to 99.97%. DL models such as EfcientNetV2, DHS-CapsNet, and CNN-based architectures such as VGG16 and GoogleNet variations have shown remarkable performance in obtaining high levels of accuracy. In addition, methods such as SSL and lightweight DL models provide encouraging outcomes in efectively managing large datasets. In general, the research emphasizes the efcacy of DL methods in successfully diagnosing cancer from histopathological pictures. It therefore indicates that DL has the potential to greatly improve medical diagnostic techniques.

Evaluating the fruit's maturity level is crucial to acquiring high-quality fruit. The skin color of some fruits may be used as one of the numerous indicators to determine whether they have achieved their peak degree of ripeness. Similar to other fruits, the skin color of an Ambarella fruit indicates its maturity. However, determining the ripeness of the Ambarella fruit was assessed manually, which is time-consuming, inefficient, taxing, requires a large number of employees, and has the potential to result in discrepancies. This study aims to classify the ripeness of the Ambarella fruit using the deep learning approach, specifically using the Convolutional Neural Network (CNN). The new family of EfficientNetV2 is trained to classify the Ambarella fruit ripeness. The pre-trained models are utilized in this work, and the training was done via transfer learning through fine-tuning. EfficientNetV2B0 achieves the highest accuracy of 100% despite having a smaller size than the other EfficientNetV2 models used in this work.

Fruits like bananas and mangoes are harvested after reaching a specific ripeness stage. Traditionally, farmers rely on manual inspection to determine ripeness, a process that can be tedious, time-consuming, expensive, and subjective. This work proposes an automatic bananas and mangoes ripeness detector utilizing computer vision technology. The detected bananas and mangoes fall into two classes: ripe and unripe. The state-of-theart YOLOv8 architecture serves as the core of the detector. Three YOLOv8 variants, YOLOv8n, YOLOv8s, and YOLOv8m, were investigated for their performance. Results show that YOLOv8s achieved the highest overall performance, 0.9991 recall, and a mean Average Precision (mAP) of 0.8897. While YOLOv8m achieved the highest precision of 0.9995, YOLOv8n is the most miniature model, making it suitable for deployment on devices with limited resources.

Introduction: Alzheimer’s disease (AD) is a progressive neurodegenerative disorder that currently affects over 55 million individuals worldwide. Conventional diagnostic approaches often rely on subjective clinical assessments and isolated biomarkers, limiting their accuracy and early-stage effectiveness. With the rising global burden of AD, there is an urgent need for objective, automated tools that enhance diagnostic precision using neuroimaging data.

Methods: This study proposes a novel diagnostic framework combining a fine-tuned VGG19 deep convolutional neural network with an eXtreme Gradient Boosting (XGBoost) classifier. The model was trained and validated on the OASIS MRI dataset (Dataset 2), which was manually balanced to ensure equitable class representation across the four AD stages. The VGG19 model was pre-trained on ImageNet and fine-tuned by unfreezing its last ten layers. Data augmentation strategies, including random rotation and zoom, were applied to improve generalization. Extracted features were classified using XGBoost, incorporating class weighting, early stopping, and adaptive learning. Model performance was evaluated using accuracy, precision, recall, F1-score, and ROC-AUC.

Results: The proposed VGG19-XGBoost model achieved a test accuracy of 99.6%, with an average precision of 1.00, a recall of 0.99, and an F1-score of 0.99 on the balanced OASIS dataset. ROC curves indicated high separability across AD stages, confirming strong discriminatory power and robustness in classification.

Discussion: The integration of deep feature extraction with ensemble learning demonstrated substantial improvement over conventional single-model approaches. The hybrid model effectively mitigated issues of class imbalance and overfitting, offering stable performance across all dementia stages. These findings suggest the method’s practical viability for clinical decision support in early AD diagnosis.

Conclusion: This study presents a high-performing, automated diagnostic tool for Alzheimer’s disease based on neuroimaging. The VGG19-XGBoost hybrid architecture demonstrates exceptional accuracy and robustness, underscoring its potential for real-world applications. Future work will focus on integrating multimodal data and validating the model on larger and more diverse populations to enhance clinical utility and generalizability.

Keywords: Alzheimer's disease, MRI, XGBoost, VGG19, classification, OASIS dataset.

Skin cancer, a severe condition, requires early and accurate detection to improvesurvival rates. This study presents a hybrid framework combining deep learning(DL) and machine learning (ML) models for enhanced skin cancer classification(SCC). Comprehensive preprocessing techniques improve data quality andgeneralization, including noise reduction, border removal, hair removal, contrastenhancement, and data augmentation. Advanced DL architectures, such asHRNet, VGG16, VGG19, ResNet50, and ResNet101, extract high-level featuresfrom skin lesion images. To refine feature quality, the outputs from the backbonenetworks are processed through a Channel Attention Block (CAB) and a Squeeze-and-Excitation Network (SENet), enhancing feature robustness anddiscrimination. These refined features are fused using Multi-Scale Feature Fusion(MSFF) to consolidate and strengthen feature representation. The final fusedfeatures are classified using various classifiers, including Random Forest (RF),Extreme Gradient Boosting (XGB), Categorical Boosting (CatBoost), AdaptiveBoosting (AdaBoost), Gradient Boosting Machine (GBM), and the DL classifierfrom the backbone. HRNet-XGB achieved the highest performance metrics,including accuracy, precision, recall, F1 score, specificity, and kappa.Explainable AI (XAI) techniques, including Grad-CAM, Grad-CAM++, andLIME, provide transparent and interpretable insights into decision-making. Thiscohesive framework advances automated skin cancer diagnosis, offeringimproved patient outcomes and reliable clinical integration.

ABSTRACTA large number of studies in the past months have proposed deep learning-based Artificial Intelligence (AI) tools for automated detection of COVID-19 using publicly available datasets of Chest X-rays (CXRs) or CT scans for training and evaluation. Most of these studies report high accuracy when classifying COVID-19 patients from normal or other commonly occurring pneumonia cases. However, these results are often obtained on cross-validation studies without an independent test set coming from a separate dataset and have biases such as the two classes to be predicted come from two completely different datasets. In this work, we investigate potential overfitting and biases in such studies by designing different experimental setups within the available public data constraints and highlight the challenges and limitations of developing deep learning models with such datasets. We propose a deep learning architecture for COVID-19 classification that combines two very popular classif...

Self-supervised visual representation learning traditionally focuses on image-level instance discrimination. Our study introduces an innovative, fine-grained dimension by integrating patch-level discrimination into these methodologies. This integration allows for the simultaneous analysis of local and global visual features, thereby enriching the quality of the learned representations. Initially, the original images undergo spatial augmentation. Subsequently, we employ a distinctive photometric patch-level augmentation, where each patch is individually augmented, independent from other patches within the same view. This approach generates a diverse training dataset with distinct color variations in each segment. The augmented images are then processed through a self-distillation learning framework, utilizing the Vision Transformer (ViT) as its backbone. The proposed method minimizes the representation distances across both image and patch levels to capture details from macro to micro perspectives. To this end, we present a simple yet effective patch-matching algorithm to find the corresponding patches across the augmented views. Thanks to the efficient structure of the patch-matching algorithm, our method reduces computational complexity compared to similar approaches. Consequently, we achieve an advanced understanding of the model without adding significant computational requirements. We have extensively pretrained our method on datasets of varied scales, such as Cifar10, ImageNet-100, and ImageNet-1K. It demonstrates superior performance over state-of-the-art self-supervised representation learning methods in image classification and downstream tasks, such as copy detection and image retrieval.

This paper illustrates how a learning-curve model can be generalized to investigate potential explanations of organizational learning. The paper examines the hypothesis that knowledge acquired through by learning by doing is embodied in an organization's technology by analyzing the amount of transfer that occurs across shifts within a plant. If knowledge becomes completely embodied in technology, transfer across shifts should be complete since both shifts use the same technology. Methods that can be used for studying intra-plant transfer of knowledge are presented. The methods are illustrated by analyzing data from a plant that began production with one shift and then added a second shift several months into the production program. Three aspects of transfer are analyzed: (1) carry forward of knowledge when the plant makes the transition from one to two shifts, (2) transfer across shifts after both shifts are operating, and (3) transfer across time. Results indicate that substant...

A strong classification model that can correctly detect abnormalities and neurological disorders in brain images is the main goal. The focus of this research is on improving the accuracy of MRI brain image categorization using residual networks (ResNet) methods. Improving the model's capacity to extract complex characteristics from MRI images and achieving more accurate classification results is the aim of using ResNet architectures. By conducting extensive experiments and validating our results, our project aims to attain top-notch performance in brain image classification tasks. The goal is to help improve medical diagnosis and treatment planning. A secondary goal of the research is to determine if deep learning approaches have any use in radiology, with the hope that this will lead to better medical image analysis pipelines. The main objective is to make it easier to identify neurological problems early on, which will enhance patient outcomes and allow for more calculated treatment decisions. Results proved that the proposed ResNet system achieves 98.8% overall accuracy with 98.6% sensitivity and 99% specificity.

Reading nonfiction texts with understanding is important to school success, yet many students struggle to do so. This randomized controlled trial extends previous research by contrasting an earlier iteration of a comprehension tutoring program (Comp) against a variant with strategies for transferring learning (Comp+Transfer). Participants were 189 fourth and fifth graders with weak reading comprehension. To evaluate their efficacy, we used commercially developed far‐transfer measures and experimenter‐made near‐ and mid‐transfer measures of reading comprehension. In contrast to controls, students in both programs significantly improved their understanding of near‐transfer passages. Additionally, students in Comp+Transfer improved performance on mid‐transfer passages. These findings suggest the value of teaching for transfer and the importance of measuring program efficacy with researcher‐made tests alongside commercial tests.

Food spoilage detection is critical in ensuring food safety and reducing waste. In this work, we offer a new neural network model, rotOrNot, intended for image analysisbased rotten food detection. Our method focuses on accurately identifying spoiled food, precisely spotting rot. The proposed method meets a vital demand for reducing food waste and guaranteeing food safety. Through several performance criteria, the proposed model was thoroughly tested and showed better performance than current approaches. rotOrNot specifically obtained training accuracy of 99.4%, validation accuracy of 96.95%, an outstanding ROC-AUC score of 99.97296%, and a test accuracy of 97.52%. Moreover, rotOrNot shows the fastest training time and exceeds the previous models in accuracy, so it is a quite effective tool for practical usage. This work offers a useful instrument for the automatic food rotting detection, therefore improving food safety procedures and maybe reducing food waste.

We present an empirical study of applying deep Convolutional Neural Networks (CNN) to the task of fashion and apparel image classification to improve meta-data enrichment of e-commerce applications. Five different CNN architectures were analyzed using clean and pre-trained models. The models were evaluated in three different tasks person detection, product and gender classification, on two small and large scale datasets.

We present an empirical study of applying deep Convolutional Neural Networks (CNN) to the task of fashion and apparel image classification to improve meta-data enrichment of e-commerce applications. Five different CNN architectures were analyzed using clean and pre-trained models. The models were evaluated in three different tasks person detection, product and gender classification, on two small and large scale datasets.

African languages are spoken by over a billion people, but are underrepresented in NLP research and development. The challenges impeding progress include the limited availability of annotated datasets, as well as a lack of understanding of the settings where current methods are effective. In this paper, we make progress towards solutions for these challenges, focusing on the task of named entity recognition (NER). We create the largest human-annotated NER dataset for 20 African languages, and we study the behavior of stateof-the-art cross-lingual transfer methods in an Africa-centric setting, demonstrating that the choice of source language significantly affects performance. We show that choosing the best transfer language improves zero-shot F1 scores by an average of 14 points across 20 languages compared to using English. Our results highlight the need for benchmark datasets and models that cover typologically-diverse African languages.

Öz This paper presents an optimized lightweight CNN model developed using a unique dataset introduced here for the first time to detect defects in manufacturing processes in a factory. The model performance was analyzed comparatively with widely used large-scale deep learning architectures such as VGG16 and ResNet50. All models were trained on the same original dataset, followed by the same approach in tuning hyperparameters such as learning rate, optimization algorithm, and data augmentation strategies. Performance analyses were conducted using fundamental metrics such as accuracy, precision, and F1 score, along with confusion matrices and randomly selected test images. Our proposed model attained high accuracy while reducing computational cost and significantly shortening training time compared to traditional architectures. The results demonstrate that the proposed CNN model achieves a competitive level of accuracy comparable to large-scale deep learning models while serving as a more suitable alternative for low-power hardware systems. Bu makale, bir fabrikadaki üretim süreçlerinde ortaya çıkan kusurların tespiti için burada ilk kez sunulan benzersiz bir veri kümesi kullanılarak geliştirilen, optimize edilmiş hafif bir CNN modelini tanıtmaktadır. Model performansı, VGG16 ve ResNet50 gibi yaygın kullanılan büyük ölçekli derin öğrenme mimarileriyle karşılaştırmalı olarak analiz edilmiştir. Tüm modeller, aynı özgün veri kümesi üzerinde, öğrenme oranı, optimizasyon algoritması ve veri artırma stratejileri gibi sabit hiperparametrelerle eğitilmiştir. Performans analizleri doğruluk, kesinlik ve F1 skoru gibi temel metriklerin yanı sıra, karmaşıklık matrisleri ve rastgele test görüntüleri üzerinden gerçekleştirilmiştir. Önerdiğimiz model, geleneksel mimarilere kıyasla daha düşük hesaplama maliyeti ve çok daha kısa eğitim süresi ile yüksek doğruluk elde etmiştir. Elde edilen sonuçlar, önerilen CNN modelinin büyük ölçekli derin öğrenme modelleriyle rekabet edebilecek düzeyde doğruluk sunarken, düşük güçlü donanıma sahip sistemler için daha uygun bir alternatif olduğunu göstermektedir.

In modern agriculture, ensuring plant health is essential for high crop yields and quality. Plant diseases pose risks to economies, communities, and the environment, making early and accurate diagnosis crucial. The internet of things (IoT) has revolutionized farming by enabling real-time crop monitoring and using drones and cameras for early disease detection. This technology helps farmers address challenges with precision and sustainability. This research proposes an ensemble learning model incorporating multi-class capsule networks (MCCN) and other pre-trained model with majority voting system is implemented to predict plant diseases and pests early. The research aims to develop a robust MCCN-based ensemble prediction model for timely disease identification. To evaluate the performance of the ensemble model, various key metrics, including accuracy, and loss value, are assessed. Furthermore, a comparative analysis is conducted, benchmarking the MCCN model against other well-known pre-trained models such as residual network-101 (ResNet101), visual geometry group-19 (VGG19), and GoogleNet. This research signifies a substantial stride towards the realization of IoT-driven precision agriculture, where advanced technology and machine learning contribute to the early detection and mitigation of plant diseases, ultimately enhancing crop yield and environmental sustainability.

Crack detection is essential for observing structural health and guaranteeing structural safety. The manual crack and other damage detection process is time-consuming and subject to surveyors’ biased judgments. The proposed Conv2D ResNet Exponential model for wall quality detection was trained with 5000 wall images, including various imperfections such as cracks, holes, efflorescence, damp patches, and spalls. The model was trained with initial weights to form the trained layers of the base model and was integrated with Xception, VGG19, DenseNet, and ResNet convolutional neural network (CNN) models to retrieve the general high-level features. A transfer deep-learning-based approach was implemented to create a custom layer of CNN models. The base model was combined with custom layers to estimate wall quality. Xception, VGG19, DenseNet, and ResNet models were fitted with different activation layers such as softplus, softsign, tanh, selu, elu, and exponential, along with transfer learn...

Through this work, we have proposed a deep CNN architecture based Unet model, to automatically segment thoracic CT scans images. At first, the database has been augmented using preprocessing algorithms such as rotations and filtering. The second step, we have proposed a deep CNN architecture based two models: Unet and Segnet . High performance scores including accuracy, F1-score have been reached. Moreover, the Unet proved to be efficient to segment different regions such as normal lungs, glass representing Covid-19 and background regions. Results quantification have been reached to 97.98% for the train and 95.49% for the validation. The segmented tested images represent an error too close to zero, in comparison to the ground truth. The proposed segmentor system shows to be reliable, achieving an IOU value equal to 92% and 87%, an MSE equal to 0.08 and 0.13 for the validation and test processing respectively. Thus, our proposed system could be a good candidate to help doctors for fast covid-19 diagnosis.

Pneumonia is a leading cause of morbidity and mortality worldwide, with chest X-rays serving as the primary imaging modality for diagnosis. While deep learning models, especially Convolutional Neural Networks (CNNs), have achieved high accuracy in pneumonia detection, their clinical adoption is hindered by two challenges: lack of interpretability and overconfident predictions. This study proposes a pneumonia detection framework that integrates explainability and probability calibration into CNN-based classification. Three architectures-Sequential CNN, ResNet50, and DenseNet121-were trained on the Kaggle chest X-ray dataset. Grad-CAM was employed to provide visual interpretability of predictions, while Platt Scaling was applied for post-hoc probability calibration. Experimental results show that the Sequential CNN achieved the best balance of accuracy (77%) and F1-score (0.78). Grad-CAM visualizations confirmed that the model focused on clinically relevant lung regions, enhancing trustworthiness. Platt Scaling reduced the Brier Score from 0.218 to 0.181, indicating improved calibration of prediction probabilities. These results demonstrate that incorporating explainability and calibration can enhance the transparency and reliability of AI-assisted pneumonia detection, making such systems more suitable for real-world clinical deployment.

Background In recent years, neuroimaging with deep learning (DL) algorithms have made remarkable advances in the diagnosis of neurodegenerative disorders. However, applying DL in different medical domains is usually challenged by lack of labeled data. To address this challenge, transfer learning (TL) has been applied to use state-of-the-art convolution neural networks pre-trained on natural images. Yet, there are differences in characteristics between medical and natural images, also image classification and targeted medical diagnosis tasks. The purpose of this study is to investigate the performance of specialized and TL in the classification of neurodegenerative disorders using 3D volumes of 18F-FDG-PET brain scans. Results Results show that TL models are suboptimal for classification of neurodegenerative disorders, especially when the objective is to separate more than two disorders. Additionally, specialized CNN model provides better interpretations of predicted diagnosis. Concl...

Background: Alzheimer's disease (AD) is a progressive neurodegenerative condition marked by cognitive decline and memory loss. Despite advancements in AI-driven neuroimaging analysis for AD detection, clinical deployment remains limited due to challenges in model interpretability and usability. Explainable AI (XAI) frameworks such as XRAI offer potential to bridge this gap by providing clinically meaningful visualizations of model decision-making. Methods: This study developed a comprehensive, clinically deployable AI system for AD severity classification using 2D brain MRI data. Three deep learning architectures MobileNet-V3 Large, EfficientNet-B4, and ResNet-50 were trained on an augmented Kaggle dataset (33,984 images across four AD severity classes). The models were evaluated on both augmented and original datasets, with integrated XRAI explainability providing region-based attribution maps. A web-based clinical interface was built using Gradio to deliver real-time predictions and visual explanations. Results: MobileNet-V3 achieved the highest accuracy (99.18% on the augmented test set; 99.47% on the original dataset), while using the fewest parameters (4.2M), confirming its efficiency and suitability for clinical use. XRAI visualizations aligned with known neuroanatomical patterns of AD progression, enhancing clinical interpretability. The web interface delivered sub-20 second inference with high classification confidence across all AD severity levels, successfully supporting real-world diagnostic workflows. Conclusion: This research presents the first systematic integration of XRAI into AD severity classification using MRI and deep learning. The MobileNet-V3-based system offers high accuracy, computational efficiency, and interpretability through a user-friendly clinical interface. These contributions demonstrate a practical pathway toward real-world adoption of explainable AI for early and accurate Alzheimer's disease detection.

Severe Acute Respiratory Syndrome Coronavirus 2 (SARS-Coronavirus-2 or SARS-CoV-2), which came into existence in 2019, is a viral pandemic that caused coronavirus disease 2019 (COVID-19) illnesses and death. Research showed that relentless efforts had been made to improve key performance indicators for detection, isolation, and early treatment. This paper used Deep Transfer Learning Model (DTL) for the classification of a real-life COVID-19 dataset of chest X-ray images in both binary (COVID-19 or Normal) and three-class (COVID-19, Viral-Pneumonia or Normal) classification scenarios. Four experiments were performed where fine-tuned VGG-16 and VGG-19 Convolutional Neural Networks (CNNs) with DTL were trained on both binary and three-class datasets that contain X-ray images. The system was trained with an X-ray image dataset for the detection of COVID-19. The fine-tuned VGG-16 and VGG-19 DTL were modelled by employing a batch size of 10 in 40 epochs, Adam optimizer for weight updates,...

With the connectivity today, the knowledge of English is mostly considered a prime prerequisite for most opportunities, be it education or professional. The proposed system uses the power of advanced natural language processing in order to facilitate communication despite linguistic differences as well as in learning personal and responsive. To mold content into the linguistic sensibilities of the mother tongue of the user, machine translation, sentiment analysis, and contextual awareness are enabled. The system tries to come up with more effective English language learning resources for diverse linguistic communities with the help of highperformance models. That is how the multilingual corpus is curated to represent the diversity of low-resource languages; thus, the learning materials will be relevant and accurate. Language-specific embeddings enable the system to identify and preserve the grammatical structures and expressions unique to each language, thereby making it a smoother transition to English. This adaptive strategy keeps the users motivated and involved in the entire process of learning a new language. Simplicity and accessibility have been designed within this system. Users with different levels of technological literacy can easily use it. The following metrics-improvement in language proficiency, user satisfaction, and engagement levels-will be deployed to appraise this system's effectiveness. Cross-Lingual NLP System One of its major strengths is language understanding at 96.6% accuracy.

India has a rich heritage of floral diversity and is well known for its medicinal plants but their identification is a major challenge in ayurvedic pharmaceutics. Hence, automated identification of medicinal flora is pivotal for sustainable pharmacognosy and botanical conservation. Muntingia calabura known for its potent therapeutic properties from high-resolution images. This research presents a novel Binary Flower Pollination Algorithm (BFPA) driven Deep SE-ResNeXt framework for the automated and robust identification of Muntingia calabura. In this study the authors proposed a framework that synergizes the hierarchical feature extraction capability of the SE-ResNeXt architecture with the optimization power of BFPA, inspired by the natural pollination behavior of flowering plants. In this framework, the SE-ResNeXt model, embedded with squeeze-and-excitation (SE) blocks, amplifies channel interdependencies and facilitates discriminative feature learning. Further, to fine-grained the identification ability BFPA is applied, which helped in selecting optimal feature subsets by simulating the global-local pollination balance through Lévy flight-based strategies. This binary optimization approach reduces redundancy in high-dimensional feature spaces, enhancing both computational efficiency and model accuracy. The BFPA-SE-ResNeXt with dense layer framework demonstrated better performance compared to conventional CNNs and other evolutionary optimization-based classifiers, achieving an identification accuracy of 97.50%, precision of 96.37%, and recall of 97%. This research underscores the effectiveness of bio-inspired hybrid intelligence systems for plant species identification and offers a scalable solution for real-world ecological applications.

Environmental sound classification(ESC) is the trending research area. ESC categorizes sounds such as dog barking, gunshots, and children playing in the surroundings. Due to overlapping sound signals, the presence of several audio sources while recording audio, and different distances from audio sources to the microphone make this problem complex. This study proposes a robust model for ESC, which can help in crime investigation systems, security warning systems, and the development of smart homes and hearing aids. Researchers have designed numerous frameworks for classifying surrounding events. Various techniques for ESC have been used in the past, but they are either computationally intensive or provide less accuracy. A hybrid model consisting of Convolutional Neural Network and Recurrent Neural Network for ESC is proposed to provide an accuracy of 99.89%, which is the highest till now, as far as we know. The model is a combination of both models; it is called CRNN. CRNN has already been used in a few past studies, but raw waveforms are used, and the accuracy attained is quite low. The publicly available Dataset UrbanSound8 K is used. Augmentation techniques are used to overcome the scarcity of datasets. The cepstral features are extracted and input to the CRNN. CRNN is encouraged due to its ability to capture spatial and temporal dependencies of environmental sound waves. Various hyperparameters, such as the number of LSTM layers, number of filters, batch size, momentum, and number of neurons in the LSTM layer, are altered to find the best value for hyperparameters for ESC. It is found that 0.5 momentum, 128 filters, 512 neurons in the LSTM layer, 256 batch size, and one LSTM layer give the highest accuracy. Another dataset, ESC-10, is used to validate the model. It is found that the proposed model provides considerable accuracy for ESC-10, even though it is lower than in the case of UrbanSound8 K. In the future, the model can be applied to different applications and datasets.

The incorporation of Artificial Intelligence (AI), especially deep learning models, into cybersecurity frameworks has greatly improved the identification and mitigation of cyber threats. Nonetheless, these smart systems encounter a significant and rising threat—adversarial attacks. Malicious entities create subtle alterations in network traffic or system actions that mislead AI models into misidentifying threats as harmless, facilitating evasion tactics that can circumvent real-time intrusion detection systems (IDS). This study investigates the susceptibility of deep learning-based Intrusion Detection Systems (IDS) to adversarial examples and suggests a robust detection framework aimed at improving resilience against these evasion tactics. The suggested system merges adversarial training, input sanitization, and resilient model architectures, including adversarial-aware Convolutional Neural Networks (CNN) and defensive autoencoders. Employing benchmark datasets like CIC-IDS2017 and UNSW-NB15, we recreate various adversarial scenarios—created using Fast Gradient Sign Method (FGSM) and Projected Gradient Descent (PGD)—to evaluate the effect on detection performance. Experimental findings indicate that conventional DL models experience a significant decline in performance when exposed to adversarial circumstances, with accuracy decreasing by more than 20% in certain instances. Conversely, our suggested framework shows a noticeable enhancement in adversarial robustness, keeping more than 91% detection accuracy during attacks and considerably lowering false positives.

Tumor is the uncontrolled growth of cancer cells in any part of the human body. Brain tumoris the leading cause of cancer deaths worldwide among adults and childrens. Early detection of brain cancers is essential. To prevent more issues, early defect detection is essential. Healthcare physicians may discover and categorize brain tumors with the use of computational intelligence-focused tools. An essential task for diagnosing tumors and choosing the right type of therapy is classifying brain tumors. Brain tumor identification and segmentation using magnetic resonance imaging (MRI) scans is now recognized as one of the most significant and difficult research areas in the world of medical image processing. The field of medical imaging has gained greatly from the use of artificial intelligence (AI) in the form of machine learning (ML) and deep learning (DL). DL has shown significant presentation, especially in the areas of brain tumor classification and segmentation. In this work, brain tumor classification for optimizing performance using hybrid recurrent neural network (RNN) classifier is presented. Different types of brain tumors are classified using a mix of RNN and inception residual neural network (ResNet). This strategy will produce improved F1-score, precision, accuracy, and recall scores.

In Sub-Saharan Africa, Professionals visually analyse the plants by looking for disease markers on the leaves to diagnose cassava infections, however, this method is extremely subjective. Automating the identification and classification of crop diseases may improve the accuracy of professional disease diagnosis and enable farmers in remote areas to monitor their crops without the assistance of experts. Algorithms for machine learning have been used in the early detection and classification of crop diseases. Motivated by the current developments in the field of Gaussian Processes, this study proposes to integrate the transfer learning approach with a deep Gaussian convolutional neural network model (DGCNN) for the detection and classification of cassava diseases. During this study, we used MobileNet V2 and VGG16 pre-trained transfer learning models and a hybrid kernel. Experiments with MobileNet V2 and a hybrid kernel revealed an accuracy of 90.11%. Also, experiments with VGG16 and a...

Image collections, if critical aspects of image content are exposed, can spur research and practical applications in many domains. Supervised machine learning may be the only feasible way to annotate very large collections, but leading approaches rely on large samples of completely and accurately annotated images. In the case of a large forensic collection, we are aiming to annotate, neither the complete annotation nor the large training samples can be feasibly produced. We, therefore, investigate ways to assist manual annotation efforts done by forensic experts. We present a method that can propose both images and areas within an image likely to contain desired classes. Evaluation of the method with human annotators showed highly accurate classification that was strongly helped by transfer learning. The segmentation precision (mAP) was improved by adding a separate class capturing background, but that did not affect the recall (mAR). Further work is needed to both increase the accuracy of segmentation and enhances prediction with additional covariates affecting decomposition. We hope this effort to be of help in other domains that require weak segmentation and have limited availability of qualified annotators.

Cancers of the skin are particularly dangerous. Skin cancer may arise due to genetic abnormalities or mutations in DNA that are not correct. Deaths have skyrocketed due to a lack of awareness about warning indicators and prevention. Hence, timely identification is imperative to avoid the dissemination of cancer. Our goal in doing this groundbreaking study was to address the pressing problem of identifying different forms of skin cancer, such as benign, malignant, normal, and acne. We began the development of an advanced and unique deep learning model, Mix Conv Dense GRU (MCD-GRU), by utilizing a carefully selected dataset that included 3932 cases of skin cancer data. Regarding validation accuracy, the model obtained 99.90% and 99.12% in training. The MCD-GRU model performed exceptionally well during both the training and validation stages. Thorough testing was used to confirm this degree of precision, and the results showed a testing accuracy of 99.13%. This method shows the greatest skin cancer detection accuracy. This article explains how deep learning might enhance skin cancer detection. This innovative approach may improve health security and safety in the future.

The manual diagnosis of diabetic retinopathy (DR) is often invasive, time-consuming, expensive, and prone to human error. Additionally, it can be subjective, depending on the clinician's professional experience. Recently, automated computer-aided diagnosis (CAD) systems have significantly reduced the time and effort required for diagnosis while achieving superior performance compared to traditional methods. Researchers have extensively explored deep learning (DL) and convolutional neural networks (CNNs) for diagnosing DR from fundus images, yielding promising results and offering a viable alternative to conventional diagnostic approaches. In this study, a hybrid model named EffNet-SVM is proposed for the classification of DR and no DR cases using retinal fundus images. The model is trained and tested using the Asia Pacific Tele-Ophthalmology Society (APTOS) dataset, which includes both DR and no DR images. The EffNet-SVM utilizes EfficientNetV2-Small for feature extraction from input fundus images, and the extracted features are then classified using a support vector machine (SVM) with a radial basis function (RBF) kernel. The EffNet-SVM model outperformed eight state-of-the-art DL models from the literature, achieving the highest accuracy of 97.26%. Performance metrics validate that the proposed hybrid model can be effectively integrated into CAD systems for the automated analysis of fundus images. INDEX TERMS Diabetic retinopathy, deep features, EfficientNet, hybrid CNN, support vector machines, transfer learning.

COVID cases and its variants is noted enormously in the past three years. In many medical cases, lung infections such as viral pneumonia, bacterial pneumonia have been initially interpreted as COVID-19. Hence, the proposed work is concentrating on differentiating these lung infection types. This work focuses on using neutrosophic approach of classifying into True (T), False (F) and Indeterminacy (I) set membership to reduce the fuzziness and retain more significant information for feature extraction of the opacity to differentiate the types of lung infections. Initially, the images are preprocessed by alpha-mean and beta-enhancement operation to reduce the indeterminacy and enhancing the image components as the range of lung opacity levels to determine the types. Then, these neutrosophic set enhanced images are fed to various deep learning models like ResNet-50, VGG-16 and XGBoost for classification. Experiments are conducted on ActualMed COVID-19 Chest X-ray and COVID-19 radiograph...

Speech synthesis, the technology that converts text into spoken words, has advanced significantly for high-resource languages like English, Spanish, and Mandarin. However, many languages spoken by millions of people are still underserved by speech synthesis systems due to insufficient data to train accurate models. An analysis of current solutions and difficulties in low-resource language speech synthesis development is presented in this paper because these languages have restricted data and linguistic resources. The paper examines different research strategies for advancing speech synthesis techniques in these languages by borrowing data from rich-resource languages to enhance models used in poor-resource systems. The primary method involves the comparison of linguistic models that operate between highly supported languages to adapt them to languages that only have limited information available. Acoustic models requiring advancement represent a crucial need to process correctly the phonetic and prosodic attributes within these languages because their traits lack proper representation in small data collections. The paper investigates different data augmentation methods, which include synthetic data generation and multilingual corpus use to enhance available training resources. These methods enhance the quality of computer-generated speech by making it sound more natural while ensuring that the words remain clear and understandable, even in low-resource languages. Notably, data augmentation techniques expanded dataset sizes by up to 27 times, improving model robustness. Additionally, models like WaveNet achieved a Mean Opinion Score (MOS) of 4.21, while Tacotron-based systems reached MOS scores as high as 4.38, demonstrating high naturalness even with limited training data. Experiments using less than three hours of training data in low-resource settings initially showed Word Error Rates (WER) exceeding 50-80%, but data augmentation and transfer learning significantly reduced these error rates. By reviewing the present state of research, this paper offers an overview of the progress made in low-resource language speech synthesis. It identifies the potential future directions for this field. Ultimately, the goal is to make speech synthesis more accessible for underrepresented languages, helping with language preservation, assistive technologies, and broader accessibility.

Due to the rapid spread of COVID-19 and its induced death worldwide, it is imperative to develop a reliable tool for the early detection of this disease. Chest X-ray is currently accepted to be one of the reliable means for such a detection purpose. However, most of the available methods utilize large training data, and there is a need for improvement in the detection accuracy due to the limited boundary segment of the acquired images for symptom identifications. In this study, a robust and efficient method based on transfer learning techniques is proposed to identify normal and COVID-19 patients by employing small training data. Transfer learning builds accurate models in a timesaving way. First, data augmentation was performed to help the network for memorization of image details. Next, five state-of-the-art transfer learning models, AlexNet, MobileNetv2, ShuffleNet, SqueezeNet, and Xception, with three optimizers, Adam, SGDM, and RMSProp, were implemented at various learning rate...

One of the primary needs of humans is food, which can be obtained through farming. Not only does agr iculture meet the necessities of humankind, but it is also a primary source of employment. Agriculture is the main driver of employment and economic growth for a growing country such as India. For a thriving agricultural and economic sector, plant disease identification is a crucial concern to enhance productivity requires attention. Conventional approaches for detecting plant diseases are gruelling, timeconsuming and demand a great deal of experience. Plant disease can be detected in a timely manner as it emerges on plant leaves, with the utilization of numerous machine learning and deep learning approaches. These high precision and time-efficient techniques are the way forward to generate qualitative farm produce. The current article examined a few methods currently used for analysing data sources, feature extraction, data augmentation and classification of plant diseases. PRISMA guidelines have been utilized to select the articles, using various keywords from peer-reviewed articles published in several databases between 2016 and 2025. Following the removal of studies based on the abstract, title, full text and conclusion, 75 publications were found and examined for their immediate relevance to plant disease recognition and categorization. Of these, 45 publications have been selected for this systematic review.

Content-Based Image Retrieval (CBIR) in computer vision applications, enables retrieval of images reflecting user intent. Traditionally CBIR is based on image processing techniques. With the emergence of Artificial Intelligence (AI), it is now possible to realize CBIR using learning-based approaches. Particularly deep learning techniques such as Convolutional Neural Network (CNN) are efficient for image analysis. In this paper, we proposed a framework known Medical Content Based Image Retrieval System (MCBIRS), which exploits pre-trained CNN variants for retrieving medical images based on image input. The framework has an offline phase for extracting visual features from training data and an online phase for processing given user queries. The descriptors obtained by CNN variants in the offline phase are persisted in a database. These are later used in the online phase to compute the distance between persisted descriptors and input image descriptor. A set of closely matching images are returned against the query image based on similarity. We proposed an algorithm known as Learning-based Medical Image Retrieval (LbMIR) to realize MCBIRS. We also implemented a re-ranking of results retrieved by the framework using other techniques. The performance of LbMIR is evaluated and compared with the state-of-the-art methods such as Bag of Visual Words (BoVW) and Histogram of Oriented Gradients (HOG). Empirical results using medical image dataset revealed that CNN variants outperformed BoVW and HOG methods. On test data, the highest performance is achieved by the proposed system with 90% mean top-k precision, demonstrating its practical implications. On the training data highest performance is achieved by proposed system (CNN variants) re-ranked with HOG with 92.30% mean top-k precision.

Learning standards (frequently referred to as academic standards, course curriculum etc.) define the specific structure of an educational program. Learning standards contain a list of instructions specifying various skills that students should learn at different points during their learning progression. For example,"calculate the area of a triangle" is one such instruction in a 6 th grade geometry curriculum. Currently these instructions are imparted using prescribed textbooks or lesson plans which have been labeled with learning standard instructions. Teachers and students use this labeled learning content to identify relevant material for teaching and studying. However with an increasing amount of users as well as publisher generated content in recent days, teachers and students may want to refer to additional content apart from prescribed textbooks for their teaching/learning needs which is not labeled with learning standard instructions. Manually identifying the appropriate learning standard instruction for each learning content is time consuming and not scalable especially since learning standards frequently contain thousands of instructions, and subject to periodic revision. In this paper, we address the problem of automatically labeling digital learning content with the learning standards. Towards this goal, we first build semantic representations of the learning standard instructions using external knowledge sources such as Wikipedia and domain text books. These semantic representations are then used in a framework which utilizes structural constraints imposed by the hierarchy of the learning standards to assign labels to the learning materials. We demonstrate the usefulness of our approach on a collection of high school learning materials that were labeled by curriculum experts from a US school district according to a publicly available learning standard. The system developed has been deployed and is in use by the school district. To the best of our knowledge we are the first to attempt this novel task and develop such a system.

Large-scale generative language models such as GPT-3 are competitive few-shot learners. While these models are known to be able to jointly represent multiple languages, their training data is dominated by English, potentially limiting their cross-lingual generalization. In this work, we train multilingual generative language models on a corpus covering a diverse set of languages, and study their few-and zeroshot learning capabilities in a wide range of tasks. Our largest model with 7.5 billion parameters sets new state of the art in few-shot learning in more than 20 representative languages, outperforming GPT-3 of comparable size in multilingual commonsense reasoning (with +7.4% absolute accuracy improvement in 0-shot settings and +9.4% in 4-shot settings) and natural language inference (+5.4% in each of 0-shot and 4-shot settings). On the FLORES-101 machine translation benchmark, our model outperforms GPT-3 counterparts on 171 out of 182 directions with 32 training examples, while surpassing the official supervised baseline in 45 directions. We conduct an in-depth analysis of different multilingual prompting approaches, showing in particular that strong in-context few-shot learning performance across languages can be achieved via cross-lingual transfer through both templates and demonstration examples. 1

Hate speech detection is complex; it relies on commonsense reasoning, knowledge of stereotypes, and an understanding of social nuance that differs from one culture to the next. It is also difficult to collect a large-scale hate speech annotated dataset. In this work, we frame this problem as a few-shot learning task, and show significant gains with decomposing the task into its "constituent" parts. In addition, we see that infusing knowledge from reasoning datasets (e.g. ATOMIC 20 20 ) improves the performance even further. Moreover, we observe that the trained models generalize to out-of-distribution datasets, showing the superiority of task decomposition and knowledge infusion compared to previously used methods. Concretely, our method outperforms the baseline by 17.83% absolute gain in the 16-shot case.

His inspiration has been stimulated us to complete our thesis. We like to express our special thanks to Professor Dr. S.M. Jahangir Alam Head of the Department of Computer Science and Engineering (CSE) for his valuable suggestions, positive advices, encouragement and sincere guidance throughout our thesis work. We also convey my special thanks and gratitude to all of the respected teachers of the department of CSE including other departments. We sincerely thank to respected hon'ble Vice Chancellor sir and respective staffs, Chief of Army and staffs, board of trustees and all members including academic members, and so on, those who led this institution and made us a bright future. We would like to thank all of our friends and the staffs of the department for their valuable suggestions and assistance. Finally, We would like to thank our parents and our family for their steady love and great support during our study.

Survivors following very premature birth (i.e., ≤ 32 weeks gestational age) remain at high risk for neurodevelopmental impairments. Recent advances in deep learning techniques have made it possible to aid the early diagnosis and prognosis of neurodevelopmental deficits. Deep learning models typically require training on large datasets, and unfortunately, large neuroimaging datasets with clinical outcome annotations are typically limited, especially in neonates. transfer learning represents an important step to solve the fundamental problem of insufficient training data in deep learning. in this work, we developed a multi-task, multi-stage deep transfer learning framework using the fusion of brain connectome and clinical data for early joint prediction of multiple abnormal neurodevelopmental (cognitive, language and motor) outcomes at 2 years corrected age in very preterm infants. the proposed framework maximizes the value of both available annotated and nonannotated data in model training by performing both supervised and unsupervised learning. We first pre-trained a deep neural network prototype in a supervised fashion using 884 older children and adult subjects, and then re-trained this prototype using 291 neonatal subjects without supervision. Finally, we fine-tuned and validated the pre-trained model using 33 preterm infants. Our proposed model identified very preterm infants at high-risk for cognitive, language, and motor deficits at 2 years corrected age with an area under the receiver operating characteristic curve of 0.86, 0.66 and 0.84, respectively. employing such a deep learning model, once externally validated, may facilitate risk stratification at term-equivalent age for early identification of long-term neurodevelopmental deficits and targeted early interventions to improve clinical outcomes in very preterm infants. Survivors following very premature birth (i.e., ≤ 32 weeks gestational age) 1,2 remain at high risk for neurodevelopmental impairments, thereby increasing their risk for poor educational, health, and social outcomes . Unfortunately, it may take 2-5 years from birth to identify such high-risk infants. Early, accurate identification, soon after birth, could pave the way for domain-specific, intensive early neuroprotective therapies during a critical window for optimal neuroplasticity. In addition to a variety of clinical prognostic factors, atypical brain connectome has been reported in children who develop adverse neurodevelopmental outcomes, suggesting that brain connectome aberrations could serve as a promising prognostic biomarker . The brain connectome is a comprehensive description of the brain's structural and functional connections in terms of brain networks . Brain connectome patterns are formed during brain maturation and reshaped in cases of prematurity or perinatal brain injury 10 . Recent advances in deep learning techniques have made it possible to

Survivors following very premature birth (i.e., ≤ 32 weeks gestational age) remain at high risk for neurodevelopmental impairments. Recent advances in deep learning techniques have made it possible to aid the early diagnosis and prognosis of neurodevelopmental deficits. Deep learning models typically require training on large datasets, and unfortunately, large neuroimaging datasets with clinical outcome annotations are typically limited, especially in neonates. Transfer learning represents an important step to solve the fundamental problem of insufficient training data in deep learning. In this work, we developed a multi-task, multi-stage deep transfer learning framework using the fusion of brain connectome and clinical data for early joint prediction of multiple abnormal neurodevelopmental (cognitive, language and motor) outcomes at 2 years corrected age in very preterm infants. The proposed framework maximizes the value of both available annotated and non-annotated data in model t...