Academia.eduAcademia.edu

Figure 4 – uploaded by Mattia Neri

See full PDFdownloadDownload figure
The three methods struggle to generalize with respect to brightness and contrast variability, yielding to generally poor segmentation results in challenging imaging conditions like the ones that characterize the initial part of our dataset (Fig. S4, 40 um). Subpar detection performance is observed even in the other parts of the dataset, with many medium and low- intensity objects being completely lost in segmentation (Fig. S4, 200 and 400 ym).

Figure 4 The three methods struggle to generalize with respect to brightness and contrast variability, yielding to generally poor segmentation results in challenging imaging conditions like the ones that characterize the initial part of our dataset (Fig. S4, 40 um). Subpar detection performance is observed even in the other parts of the dataset, with many medium and low- intensity objects being completely lost in segmentation (Fig. S4, 200 and 400 ym).

Related Figures (12)

Fig. S1. Picture of various samples before (PBS) and after (TDE) the clearing protocol. Acronym: HME (Hemimegalencephaly), FCD (Focal cortical dysplasia). SWITCH/TDE sample clearing. To demonstrate the widespread applicability of the SWITCH/TDE protocol different samples were cleared. Supplementary figure S1 shows 4 autoptic pieces from healthy subjects and 5 blocks from diseased patients of various ages, kept in formalin for different amounts of time before and after the clearing procedure.
Fig. S1. Picture of various samples before (PBS) and after (TDE) the clearing protocol. Acronym: HME (Hemimegalencephaly), FCD (Focal cortical dysplasia). SWITCH/TDE sample clearing. To demonstrate the widespread applicability of the SWITCH/TDE protocol different samples were cleared. Supplementary figure S1 shows 4 autoptic pieces from healthy subjects and 5 blocks from diseased patients of various ages, kept in formalin for different amounts of time before and after the clearing procedure.
Fig. S2. Representative two-photon images of tissue stained with anti-NeuN antibody (in red) and DAPI (in green) at different temperatures and incubation times. Scale bar = 50 um. Background (S/B) analysis was performed with Fiji (http://fiji.sc/Fiji) to assess the best contrast: mean intensity of 10 square of 25px were analyzed for each condition, media and standard deviation were then calculated using OriginPro 9.0 (OriginLab Corporation). The protocol with the highest signal amplification corresponds to 4°C of incubation temperature and 24 hours of incubation time of the primary antibody.
Fig. S2. Representative two-photon images of tissue stained with anti-NeuN antibody (in red) and DAPI (in green) at different temperatures and incubation times. Scale bar = 50 um. Background (S/B) analysis was performed with Fiji (http://fiji.sc/Fiji) to assess the best contrast: mean intensity of 10 square of 25px were analyzed for each condition, media and standard deviation were then calculated using OriginPro 9.0 (OriginLab Corporation). The protocol with the highest signal amplification corresponds to 4°C of incubation temperature and 24 hours of incubation time of the primary antibody.
Fig. S3. First row: a representative middle plane (~ 200 um) of the mesoscopic reconstruction of the four analyzed human brain specimens obtained with TPFM. Scale bar = 1mm. Bottom rows: magnified insets of each specimen showing the native resolution of the acquisition. Images were selected from the top, middle plane and bottom of the stacks indicated with a white box in the first row. Scale bar = 50 tm.
Fig. S3. First row: a representative middle plane (~ 200 um) of the mesoscopic reconstruction of the four analyzed human brain specimens obtained with TPFM. Scale bar = 1mm. Bottom rows: magnified insets of each specimen showing the native resolution of the acquisition. Images were selected from the top, middle plane and bottom of the stacks indicated with a white box in the first row. Scale bar = 50 tm.
Fig. S5. The CNN architecture. Block scheme of the architecture of the CNN with 3 convolutional layers and 3 Fully Connected layers.
Fig. S5. The CNN architecture. Block scheme of the architecture of the CNN with 3 convolutional layers and 3 Fully Connected layers.
Fig. S6. (a) Neuronal segmentation workflow of the 2.5D approach. (b) A representative image undergoing the CNN analysis. From the native image to cells contour segmentation. Scale bar = 100 pm
Fig. S6. (a) Neuronal segmentation workflow of the 2.5D approach. (b) A representative image undergoing the CNN analysis. From the native image to cells contour segmentation. Scale bar = 100 pm
The network parameters are as follows: Optimizer hyper-parameters are as follows:
The network parameters are as follows: Optimizer hyper-parameters are as follows:
Fig. S7. 3D representation of the neurons of a stack manually annotated by the operator (in blue), automatically identified by the 2.5D approach (in red), and the superposition of the two. 2.5D statistical assessment. The statistical assessment of the 2.5D performance wa: determined by analyzing four representative stacks of 100 x 100 x 100 um’, one for eack specimen and a representative stack of 100 x 100 x 450 um’ to determine the accuracy along the full depth. Each stack was independently manually annotated by an operator anc automatically segmented by the 2.5D approach. The manual segmentation has_ beer performed on LAIRA® (https://laira.bioretics.com) and stored in the Ximage open-source format (https://github.com/bioretics/ximage). LAIRA® is available free of charge in tria mode. Supplementary Figure S7 shows the comparison between the manual annotations anc the automatic reconstruction for one of these sub-volumes.
Fig. S7. 3D representation of the neurons of a stack manually annotated by the operator (in blue), automatically identified by the 2.5D approach (in red), and the superposition of the two. 2.5D statistical assessment. The statistical assessment of the 2.5D performance wa: determined by analyzing four representative stacks of 100 x 100 x 100 um’, one for eack specimen and a representative stack of 100 x 100 x 450 um’ to determine the accuracy along the full depth. Each stack was independently manually annotated by an operator anc automatically segmented by the 2.5D approach. The manual segmentation has_ beer performed on LAIRA® (https://laira.bioretics.com) and stored in the Ximage open-source format (https://github.com/bioretics/ximage). LAIRA® is available free of charge in tria mode. Supplementary Figure S7 shows the comparison between the manual annotations anc the automatic reconstruction for one of these sub-volumes.
Fig. S8. The figure shows the 10 selected lines used for the analysis for specimens 1, 2, and 3 drown on top of the maps of figure 5d. Depth accuracy quantification. To describe the variability of the 2.5D along with depth, we plotted the z-profiles across the cortex along 10 different lines that were drawn orthogonally to the cortical layers in specimens 1, 2, and 3 using Fiji. To avoid the distribution artifacts introduced by the imaging (e.g alteration of the surface and the bottom of the sample due to the fact that the tissue is not completely flat), for each sample we selected the middle planes (6 depth=200um) that present intact surfaces. Supplementary Figures S8 shows for each specimen the 10 selected lines. Then, we calculated and plotted the mean value and the one standard deviation of each profile along Z (Supplementary Figures S9, S10, $11). The stability of the profile along with the cortical depth and the stack depth proves that the 2.5D evaluation is stable between different samples, in various regions of the cortex, and along with the stack depth.
Fig. S8. The figure shows the 10 selected lines used for the analysis for specimens 1, 2, and 3 drown on top of the maps of figure 5d. Depth accuracy quantification. To describe the variability of the 2.5D along with depth, we plotted the z-profiles across the cortex along 10 different lines that were drawn orthogonally to the cortical layers in specimens 1, 2, and 3 using Fiji. To avoid the distribution artifacts introduced by the imaging (e.g alteration of the surface and the bottom of the sample due to the fact that the tissue is not completely flat), for each sample we selected the middle planes (6 depth=200um) that present intact surfaces. Supplementary Figures S8 shows for each specimen the 10 selected lines. Then, we calculated and plotted the mean value and the one standard deviation of each profile along Z (Supplementary Figures S9, S10, $11). The stability of the profile along with the cortical depth and the stack depth proves that the 2.5D evaluation is stable between different samples, in various regions of the cortex, and along with the stack depth.
Fig. S9. Plots of cortical depth profiles (i.e. along Z) are shown together with the mean value and the one standard deviation for each of the 10 lines of specimen 1 showed in Fig $7.
Fig. S9. Plots of cortical depth profiles (i.e. along Z) are shown together with the mean value and the one standard deviation for each of the 10 lines of specimen 1 showed in Fig $7.
Fig. S10. Plots of cortical depth profiles (i.e. along Z) are shown together with the mean value and the one standard deviation for each of the 10 lines of specimen 2 showed in Fig S7
Fig. S10. Plots of cortical depth profiles (i.e. along Z) are shown together with the mean value and the one standard deviation for each of the 10 lines of specimen 2 showed in Fig S7
Fig. S11. Plots of cortical depth profiles (i.e. along Z) are shown together with the mean value and the one standard deviation for each of the 10 lines of specimen 3 showed in Fig S7
Fig. S11. Plots of cortical depth profiles (i.e. along Z) are shown together with the mean value and the one standard deviation for each of the 10 lines of specimen 3 showed in Fig S7
Fig. S12. Layer masks (L1 red, L2 white, L3 blue, L4 green, L5 light blue, L6 yellow) of the middle plane of samples 1,2, and 3. Grey matter mask (in grey) is shown for sample 4. Manual segmentation of the grey matter. In order to identify the different layers of the grey matter of the cortex, we manually segmented the cell and mean volume distribution maps using the software Fiji (Supplementary figure S12). Big blood vessels, tissue holes/breakages, and imaging artifacts were not reckoned drawing the masks. Sample 4 shows a disruption of the structural organization of the cortex, making layer classification impossible, only grey matter were segmented. The total volume for each mask was obtained performing this operation for all the z-plane of the sample.
Fig. S12. Layer masks (L1 red, L2 white, L3 blue, L4 green, L5 light blue, L6 yellow) of the middle plane of samples 1,2, and 3. Grey matter mask (in grey) is shown for sample 4. Manual segmentation of the grey matter. In order to identify the different layers of the grey matter of the cortex, we manually segmented the cell and mean volume distribution maps using the software Fiji (Supplementary figure S12). Big blood vessels, tissue holes/breakages, and imaging artifacts were not reckoned drawing the masks. Sample 4 shows a disruption of the structural organization of the cortex, making layer classification impossible, only grey matter were segmented. The total volume for each mask was obtained performing this operation for all the z-plane of the sample.
580 California St., Suite 400
San Francisco, CA, 94104
© 2025 Academia. All rights reserved