Semi-supervised medical image lesion detection based on multi-head feature fusion

Abstract

Semi-supervised object detection methods hold significant application value in medical image lesion detection. However, due to two major challenges—the high complexity of medical images and the sparsity of lesion instances—existing methods often struggle to achieve satisfactory performance. To address these issues, this paper proposes a Medical Image Semi-Supervised Object Detection network, termed MISSOD. The framework incorporates a Multi-head Feature Fusion Module (MFFM), which alleviates feature dispersion caused by high image complexity through parallel multi-scale feature aggregation and a channel attention mechanism. Additionally, a confidence-based IoU regression loss is introduced to continuously optimize the regression error between pseudo-labels and predicted bounding boxes throughout the training process, thereby improving the utilization of sparse lesion instances. Evaluated on three medical image lesion detection datasets under varying labeled data ratios (i.e., 100%, 10%, 5%, 2%, and 1%), MISSOD achieves consistent improvements in metrics such as AP and Recall, demonstrating its effectiveness. Thus, under realistic medical imaging data constraints, the proposed method offers an effective solution for semi-supervised detection in complex medical imaging scenarios. All source code is publicly available at https://github.com/hexiang-zhang/MISSOD.