LCA-Med: A lightweight cross-modal adaptive feature processing module for detecting imbalanced medical image distribution.

Abstract

Data distribution discrepancy across datasets is one of the major obstacles hindering the improvement of the accuracy of cross-domain adaptive detection of medical images. To address this challenge, we propose a novel lightweight cross-modal adaptive detection module named LCA-Med (LCaM). The proposed module boasts a lightweight structure and a minimalistic parameter count, thereby facilitating its integration into the anterior segment of a diverse array of foundational and downstream networks. It is adept at serving as a feature preprocessor, proficiently extracting pertinent information regrading pathologies from a array of images (image modality) produced through varied medical imaging techniques, all guided by the input of prompts (text modality). We also propose a novel cross-modal medical image adaptive detection method, LCA-Med CNX (LCaM-CNX), and a novel cross-domain adaptive detection training paradigm that incorporates generated dataset groups, an attention module, and a meta-heuristic algorithm. Experimental results on six medical image datasets compared with ten state-of-the-art methods demonstrate that the LCaM-CNX trained following the proposed paradigm achieves the best performance on five datasets and competitive performance on the other dataset. Notably, our method outperforms the state-of-the-art methods more when the data distribution is more imbalanced.