OSNet-DCN: Integrating Deformable Feature Learning for Tracking Lesion in Endoscopic Videos

As artificial intelligence continues to permeate diverse fields, particularly healthcare, deploying deep neural networks in the analysis and processing of medical images has emerged as a cornerstone of advanced diagnostics. Although numerous methods for tracking the lesion have been developed in endoscopic videos, existing studies focus mainly on polyps with relatively distinct morphologies. In this research, we target lesions in the upper gastrointestinal tract, addressing the primary challenge of high tissue and mucosal deformation coupled with the complex motion of the endoscopic camera. To address this, we propose an improvement to the OSNet feature extractor within the StrongSORT pipeline by incorporating a Deformable Convolution Network (DCN), which excels at capturing complex geometric features corresponding to the high deformation of target objects. Experimental results demonstrate that incorporating DCN into the final layers of OSNet significantly improves performance, achieving a mean Average Precision (mAP) of 0.869 and a Rank-1 Accuracy of 0.954. This approach showcases robust tracking capabilities under challenging conditions, paving the way for more accurate and reliable lesion tracking in clinical endoscopy. The code is available at https://github.com/shredder0812/osnet_dcn.