FMC-Net: Fine-grained multi-lesion classification in wireless capsule endoscopy via attention-guided feature interaction.

Accurate detection of multiple lesions in wireless capsule endoscopy (WCE) images is crucial for gastrointestinal disease diagnosis. However, high inter-class similarity, intra-class variability, and the frequent coexistence of different lesion types within one image make this task highly challenging. To overcome these limitations, we propose a fine-grained multi-lesion classification network (FMC-Net) that enhances lesion discrimination and support single-label/multi-label classification through an innovative attention-guided feature selection module and a flexible classifier. The attention-guided feature selection module adopts two independent branches, one is designed as a weakly supervised selector to select local features, another is built with path augmentation blocks to achieve fused global features. After that, a global-local attention module is implemented to enhance the interaction between global and local features, thus improving the model' s fine-grained recognition and discrimination capability of lesions. Additionally, a flexible classifier with an interchangeable activation function further supports both single-label and multi-label learning. Extensive experiments on private WCE datasets demonstrate that FMC-Net outperforms state-of-the-art methods for sorting different types of images, including normal, bleeding, ulcer, erosion and polyp. In single-label classification, FMC-Net improves accuracy by over 2.68% on independent test dataset and over 5.15% on complete WCE cases. In multi-label classification, FMC-Net achieves an average accuracy of 81.78%, exceeding existing methods by at least 6.29%. The promising results demonstrate that the proposed method achieves competitive and robust detection performance in complex WCE scenarios, highlighting its potential for clinical application in automated gastrointestinal lesion detection and contributing to more accurate and efficient diagnostic workflows.