Counterfactual Explainable Gastrointestinal and Colonoscopy Image Segmentation

Abstract

Segmenting medical images accurately and reliably is crucial for disease diagnosis and treatment. Due to the wide assortment of objects’ sizes, shapes, and scanning modalities, it has become more challenging. Many convolutional neural networks (CNN) have recently been designed for segmentation tasks and achieved great success. This paper presents an optimized deep learning solution using DeepLabv3+ with ResNet-101 as its backbone. The proposed approach allows capturing variabilities of diverse objects. It provides improved and reliable quantitative and qualitative results in comparison to other state-of-the-art (SOTA) methods on two publicly available gastrointestinal and colonoscopy datasets. Few studies show the inadequacy of stable performance in varying object segmentation tasks, notwithstanding the sizes of objects. Our method has stable performance in the segmentation of large and small medical objects. The explainability of our robust model with benchmarking on SOTA approaches for both datasets will be fruitful for further research on biomedical image segmentation.