Fig. 3

Flow diagram of the pixel-based approach for classification and semantic segmentation of MP and non-MP regions from bladder H&E-stained histopathological images. In the model training step, the annotated images are stain normalized. The stain normalized images and their corresponding ground truth masks are split into patches. The semantic segmentation architecture is trained on that dataset to classify each pixel as MP or non-MP region. In the patch-based inference step, the test image is stain normalized and split into patches. Each patch is passed through the trained semantic segmentation architecture to obtain a similar patch where each pixel of the patch is assigned a probability for MP and non-MP. All the predicted patches are merged together to form a segmented test image. Thus, the predicted output image contains either white pixels (MP) or black pixels (non-MP). This output image is further post-processed to obtain a binary image with smooth edges and minimal noisy pixels. In the whole image-based inference step, the test image is stain normalized and directly passed through the trained semantic segmentation architecture to obtain a segmented test image. The predicted output image contains either white pixels (MP) or black pixels (non-MP). The output image is post-processed similar to the patch-based inference step