Mamba-DDPM-BSA: Diffusion model based boundary sampling algorithm for imbalanced classification

Abstract

Data category imbalance is one of the major challenges in the field of medical image classification. This imbalance seriously affects the accuracy and reliability of the classification model, posing potential risks to doctors’ diagnosis and treatment. This paper proposes a Mamba-DDPM-BSA method to address the imbalanced classification issue of medical image. Firstly, the generative model Mamba-DDPM is designed for the synthesis of medical image samples. It utilizes Mamba’s global modeling capability and linear computational efficiency to improve the quality of generated samples by improving DDPM (Denoising Diffusion Probabilistic Model). Secondly, by oversampling training samples in boundary regions, the proposed Boundary Sampling Algorithm (BSA) enables synthesizer focuses more on decision boundary areas when fitting sample distributions. This approach generates more samples near the decision boundary, pushing the decision boundary that originally intrudes into the minority class distribution towards the true distribution. Finally, a Mamba-DDPM-BSA method is proposed, which adopts an interactive synthesis method and makes full use of diffusion generation model and Boundary Sampling Algorithm to interact with the classification model, aiming to synthesize images that target the defects of the classification model to improve the discriminative ability and robustness of the classifier. Experiments based on HAM10000 data set show that Mamba-DDPM-BSA reaches 81.03%, 82.14%, and 82.71% on Matthew’s correlation coefficient, Balanced Accuracy, and Macro F1, respectively. The proposed method is superior to the traditional imbalanced classification method.