A Generalized Contrast-Adjustment Guided Growth Method for Medical Image Segmentation

Qian Tang, Qikui Zhu, Yongchao Xu, Bo Du

Abstract: Data augmentation methods have proven effective in addressing the scarcity of training data. However, two unsolved challenges still limit their potential in medical image segmentation. Specifically, (1) existing data augmentation methods cannot establish the interaction between two tasks. They treat data augmentation and segmentation as two independent tasks, which ignores the inter-correlation. (2) They cannot enhance the weak boundary of utmost significance in medical image segmentation. Instead, they focus exclusively on increasing the number and diversity of training samples. We propose a novel and generalized contrast-adjustment guided growth method (CaGM) with two innovations to solve the above challenges. Specifically, (1) for the first time, the concept of growth method is proposed, which innovatively unifies the segmentation and augmentation into one framework, and establishes the inter-correlation between two tasks for boosting the segmentation. (2) A novel contrast-adjustment method is proposed, which enables boosting the contrast of boundaries and tackling the challenge of weak boundaries. Experimental results on two datasets and four baseline segmentation methods demonstrate that the CaGM has exhibited remarkable generality. The CaGM significantly improves segmentation performance and outperforms state-of-the-art data augmentation methods.