SegFormer-Based Approach for Semantic Segmentation of Necrotic Tissues from Histopathology Images

The treatment plan assigned to a cancer patient will be decided by various factors and prognostic score is one among them. For Osteosarcoma and Renal Cell Carcinoma, such a decision factor is the amount of tumor necrosis created due to Neoadjuvant Chemotherapy. The main objective of this study is to evaluate the patient body's therapeutic response to Neoadjuvant Chemotherapy by quantifying the tumor necrosis accumulated in the body. The SegFormer model used in this study has efficiently portray the necrotic tissue area from the input histopathology image. It is one of the Transformer-based framework that can be applied for semantic segmentation of images. The encoder side of this model is a Transformer and its decoder side is a Multi-layer Perceptron that causes high performance with lesser computation power. Dice loss has been used to measure the loss and Intersection over Union score for the accuracy. To facilitate this research, a dataset comprising 900 images and their corresponding masks was meticulously curated with the assistance of an experienced pathologist.