Dual branch multi-scale feature fusion method for liver tumor detection

Liver tumor detection faces challenges such as complex image backgrounds, indistinct tumor boundaries, and inadequate multi-scale feature representation in existing models. To address these issues, this study proposes a dual-branch multi-scale feature fusion model named YCMFL for liver tumor detection. Firstly, a dual-branch feature extraction network is developed, integrating the strengths of both YOLOv8s and CenterNet backbone networks. The parallel dual-branch architecture enables efficient capture of multi-scale features of tumors. Secondly, the C2f_CoT module is introduced, which combines contextual attention mechanisms with convolutional operations to enhance the model’s ability to integrate global and local features during the feature fusion process. Additionally, the model incorporates a small object detection head, specifically designed to improve detection capabilities for small lesions, thereby effectively alleviating the performance bottleneck of traditional models in detecting small targets. Experimental results demonstrate that the proposed method achieves an mAP of 92.4% on the LiTS2017 dataset and 92.6% on the LT3DM dataset, representing improvements of 1.8% and 2% compared to the original models, respectively. Compared with current mainstream detection models, the proposed method exhibits superior detection performance and provides an effective solution for liver tumor detection. • Designing a Dual-Branch Feature Extraction Network to Address the Limitations of Single-Branch Networks in Feature Extraction. • Introducing the CoT Attention Mechanism into the C2f Module to Enhance Global Information Capture and Feature Fusion. • Incorporating a Small Target Detection Head to Improve Detection Accuracy for Small-Scale Targets.