Weighted U-NET++ and 2D-HMM Ensemble for Gastrointestinal Image Segmentation

Jairo Enrique Ramírez-Sánchez, Pedro A. Martínez-Barrón, Hannia Medina-Aguilar, Romeo Sánchez-Nigenda


One of the most widely used treatments for cancer of the gastrointestinal (GI) tract is radiotherapy, which requires manual segmentation of the affected organs to deliver radiation without affecting healthy cells. Deep learning techniques have been used, especially variants of U-Net, to automate the organ segmentation process, increasing the efficiency of medical treatment. However, the effective segmentation of the GI tract organs remains an open research problem due to their high capacity to deform because of body movement and respiratory function. This work proposes a methodology that develops a weighted ensemble integrating U-Net++ models and Hidden Markov Models (2D-HMM) for semantic segmentation of the stomach and bowels. Our empirical evaluation reports a score of 0.811 for the Dice coefficient using Leave-One-Out Cross-Validation, which provides robustness to the results.


Image segmentation, U-NET architecture, machine learning, hidden markov models

Full Text: PDF