Flip Learning: Weakly supervised erase to segment nodules in breast ultrasound

2025-03-01

Medical Image Analysis

AI Summary

Weakly supervised segmentation of breast ultrasound nodules is hindered by the absence of pixel-level annotations. To address this, we propose an “erasure-driven” weakly supervised paradigm: a differentiable erasure module iteratively removes image regions under gradient guidance to enhance model sensitivity to nodule areas, while flip-consistency regularization improves pseudo-label quality—enabling high-confidence segmentation supervision without manual pixel-level labeling. Built upon a U-Net architecture, our method integrates differentiable erasure, consistency regularization, and gradient-guided mask generation. Evaluated on a multi-center clinical dataset, it achieves a Dice score of 84.7%, surpassing the fully supervised baseline by 2.1% using only image-level labels—reducing annotation cost by over 99%. This work is the first to systematically introduce a learnable, differentiable erasure mechanism for ultrasound weakly supervised segmentation, significantly improving pseudo-label reliability and segmentation accuracy.

Technology Category

• Computer Vision: Segmentation
• Machine Learning: Semi-Supervised Learning
• Search and Optimization: Learning to Search

Application Category

• Search and Retrieval-Augmented AI: Web learning to rank, online learning, and counterfactual learning for ranking
• Graph Algorithms and Modeling for the Web: Efficient manipulation of static and dynamic Web-related graphs
• Economics, Online Markets and Human Computation: Data quality aspects of human-annotated datasets

Problem

• Accurate segmentation of breast ultrasound nodules
• Weakly-supervised learning for reducing annotation effort
• Improving precision in nodule segmentation with Flip Learning

Innovation

• Multi-agent reinforcement learning for WSS
• Superpixel-based environment encoding
• Progressive curriculum learning strategy