Pixel by Pixel Semantic Segmentation Approach on WSI Images for Gastric Gland Segmentation and Gastric Cancer Grade Classification Using MLP-XAI Model

Abstract

Gastric cancer remains one of the most prevailing cancers with high mortality. Timely and quantitative diagnosis stays challenging with the pathologists. H&E stain provides a color composition which distinguishes individual components of gastric histopathology images. The human eye is able to distinguish each component but fails to quantify and varies with the pathologists' opinions. The gastric histopathology components like lamina propria sometimes contain hyperchromatic nuclei and lymphocytes. This characteristic sometimes makes the diagnosis confusing as the system might incorrectly identify it as malignancy. Automation of this diagnosis is extremely crucial but can be a strong support system in gastric cancer diagnosis. This study developed a combinational neural network approach based on DeepLabV3+ and U-Net architectures. A pixel-by-pixel semantic segmentation approach is implemented to segment gland texture from gastric histopathology WSI images. A sliding window approach is employed to process the whole slide images. Various categories of gastric abnormalities classification models are implemented using Multilayer Perceptron (MLP). To interpret the classification model, the XAI technique is used, utilizing SHapley Additive exPlanations (SHAP). The model is able to categorize gastric lesions into five classes: benign, mild dysplasia, dysplasia, high-grade dysplasia, and malignant using the features nuclear-cytoplasmic ratio, GLCM, and intensity metrics. The segmentation model scored an accuracy of 96.983%, precision of 94.057%, recall of 93.835%, and F1 score of 95.497%, and the classification model achieved an accuracy of 90.36%. A framework is designed to support pathologists in making early decisions on gastric cancer.