Fig 1.
Comprehensive assessment of crypt microstructural changes during tissue injury and recovery a. standard of protocols to evaluate WSIs using software assistant semi-quantitative assessment. b. Deep learning augmented computational workflow to process digitalized WSIs across whole tissue fields with various localization.
Fig 2.
A training workflow for histopathology lab.
The decision make step is shown in text with question marker. The arrows indicate the direction. After making the basic pre-trained model, the training can be implemented using quick annotator or equivalent.
Fig 3.
a. Architecture of Multichannel Convolutional Networks used for Crypt Segmentation. The boxes represent cross-sections of square feature maps. Number of map size dimensions were annotated on the lower left, and the number of channels were labeled on the top of the correponding box. The leftmost input were 512×512 image tiles sampled from the whole slide. The rightmost of the output is the CNN’s binary ring mask prediction. Arrows represent operations and convolutions, b. Architecture of Autoencoder Network includes 4 steps. Step 1: the leftmost map represents gland patches generated from Convolutional U-Net. Step 2: Random extraction of the 24x24 patch from central and peripheral regions of the gland patch. Step 3: multiple channel convolutional autoencoder architecture. Step 4: t-SNE visualization using data from bottom neck.
Fig 4.
Segmentation results from the mouse colon tissue dataset.
Whole image of a. Control colon b. colon from DSS model, c. colon from treated DSS model. Details of b. Control colon c. colon from DSS model, d. colon from Sulfasalazine SSZ treated DSS model. For each panel, upper left: Original H&E image. Upper right: Predicted probability mask of gland (pixel level). Lower left: H&E overlaid with crypts binary mask. Lower right: H&E overlaid with predicted masks of gland (blue) and mucosa (orange red).
Fig 5.
Visualization of a dense representation of all the crypts in the study.
The convolutional autoencoder is built to learn representative of segmented glands patches obtained from U-Net. A t-SNE was shown with the typical gland image patches to exemplify the difference. Each data point represents one complete crypt image patch. Total 39655 segmented image of crypts were used in the map.
Fig 6.
Summary of image features for histology assessment at slide level.
A total number of 186 slides were used. a. Correlation map for texture features including histopathologic derived epithelium changes score by pathologist, entropy in masked gland, distances between the nearest glands, and the angle difference between the two nearest glands, GLCM (ASM, homogeneity, entropy, energy, dissimilarity, correlation, glcm_contrast). Agreement between histopathologic derived epithelium change scores and segmentation-derived entropy measures. b. Correlation scatter plot of GLCM contrast with epithelim change score. Each dot represents one individual slide. c. Correlation scatter plot of GLCM dissimilarity with epithelial change score. Each dot represents one individual slide. Pearson correlation coefficient and p-value were shown.
Fig 7.
Group comparison among Ctrl (Naïve animal), DSS (DSS-induced colitis), and Treatment (colitis treated with SSZ) at individual slide level using a. histopathologic derived epithelium change score; b. gland and mucosa ratio score; c. GLCM contrast value in masked crypts; d. GLCM dissimilarity value in masked crypts. Data were summarized for individual slides. Each data point represents the average value extracted from all segmented crypts from each slide. All statistical analyses between groups were performed using unpaired two-sided Student’s t tests with unequal variance.
Fig 8.
Group comparison among Ctrl (Naïve animal), DSS (DSS-induced colitis), and Treatment (colitis treated with SSZ) at individual subject level.
Data were summarized for individual animals. Each sample represents the average value from all slides derived from individual animal in the study. The group information was shown with three colors. a. epithelial change scored by histopathologist, b. ratio of crypt/mucosa area c. distance between two nearest crypts, d. angle changes between two nearest crypts. e: entropy from patches without applying segmentation masks; entropy from unsegmented slides, f: entropy from masked crypt, g-k: GLCM ASM, Homogeneity, energy, dissimilarity, contrast) l: canny.
Fig 9.
Integration with pathology score.
a. PCA plot of crypt image features generated from control, DSS, and SZZ groups were shown. Each dot represents one animal in the study. b. Clusters of samples from Ctrl (Naïve animal), DSS (DSS-induced colitis), and Treatment groups were shown in heatmap using of GLCM features and histopathological scores including epithelium changes and inflammation measurement. Z score normalization was performed.