Partial Labeled Gastric Tumor Segmentation via patch-based Reiterative Learning

TL;DR

This paper introduces a reiterative learning framework for gastric tumor segmentation that effectively trains on partially annotated images, achieving high accuracy without extensive manual labeling or pre-training.

Contribution

The proposed method enables effective training on partial annotations using a patch-based, reiterative approach with an overlapped region forecast algorithm, eliminating boundary errors.

Findings

Achieved a mean IOU of 0.883 on partial labeled data.

Attained a mean accuracy of 91.09%.

Won the 2017 China Big Data & AI Innovation Competition.

Abstract

Gastric cancer is the second leading cause of cancer-related deaths worldwide, and the major hurdle in biomedical image analysis is the determination of the cancer extent. This assignment has high clinical relevance and would generally require vast microscopic assessment by pathologists. Recent advances in deep learning have produced inspiring results on biomedical image segmentation, while its outcome is reliant on comprehensive annotation. This requires plenty of labor costs, for the ground truth must be annotated meticulously by pathologists. In this paper, a reiterative learning framework was presented to train our network on partial annotated biomedical images, and superior performance was achieved without any pre-trained or further manual annotation. We eliminate the boundary error of patch-based model through our overlapped region forecast algorithm. Through these advisable methods, a mean intersection over union coefficient (IOU) of 0.883 and mean accuracy of 91.09% on the partial labeled dataset was achieved, which made us win the 2017 China Big Data & Artificial Intelligence Innovation and Entrepreneurship Competitions.