Cost-Sensitive Distribution Alignment for Improving Automated Medical Diagnosis Using Retinal Fundus Photography

Domain adaptation (DA) has emerged as a promising approach to address the domain shift problem in deep learning for automated medical diagnosis. However, current approaches often overlook the imbalanced nature of different categories and primarily focus on aligning the distributions of the source and target domains globally. This oversight leads to suboptimal performance on imbalanced target datasets, as the alignment process becomes dominated by the majority class during adaptation. To tackle this limitation, this paper proposes a novel domain adaptation method called cost-sensitive distribution alignment (CSDA), which aims to enhance deep learning performance on target data. The proposed approach involves collecting a limited-sized dataset from the target domain and utilising CSDA to bridge the domain gap between the source and target domains. Specifically, to address class imbalance, cost-sensitive learning is incorporated into the distribution alignment process, giving more emphasis to the misalignment cost associated with minority class samples. More importantly, CSDA aligns the cross-domain projection distribution in the feature space with the ideal geometric distribution derived from the ground-truth labels. Unlike existing methods that directly minimise or employ adversarial learning to reduce the distribution discrepancy between the source and target domains, our proposed CSDA method focuses on minimising the semantic relationship misalignment among cross-domain samples. Experimental results on the detection of high myopia and myopic macular degeneration (MMD) demonstrate the superiority of CSDA over state-of-the-art methods. The results provide empirical evidence of CSDA’s efficacy in enhancing automated medical diagnosis.