SMGF²-YOLACT: a Spatial response, Mask vector Generation-Focus-Fusion instance segmentation for head-neck lymph nodes

Head-neck lymph nodes are important pathways for the metastasis of cervical malignant tumors. Accurate segmentation of the lesion is of great significance for clinical diagnosis and treatment evaluation. There are some problems in anchor boxes-based instance segmentation caused by differences in lesion size and vague boundary, such as redundant anchor boxes, high computational cost, weak coupling ability between instance masks and spatial positions, and limited perception ability. The SMGF²-YOLACT model with "Heatmap better than Box Mask" is proposed. The model revolves around 4 questions "How to use the hot zones of heatmap to guide the instance mask vector?", "How to generate the instance mask vector?","How to focus the instance mask vector?" and "How to fuse instance mask vectors?". Firstly, the model introduces spatial heatmap to replace anchor boxes for location, enhancing the context awareness ability; Secondly, in the heatmap high-response region, candidate mask vectors are generated based on the heatmap guidance mechanism. The candidate mask vectors are optimized by using Mask Regression and Semantic Aggregation mechanisms to enhance the consistency between instance masks and instance object; Thirdly, Top-K and Top-L focusing strategies are designed to compress the redundant mask vector set and focus on the boundary region; Finally, the model fuses the mask vector and the prototype mask to obtain semantically consistent and clear boundary instance mask. The validity of the model is verified on the head-neck lymph nodes dataset, with mAP_det and mAP_seg reaching 40.2% and 43.4% respectively, and the inference speed reaching 36FPS. The generalization ability of the model is evaluated on the BraTS brain tumor dataset, with mAP_det and mAP_seg reaching 65.0% and 64.3% respectively, and the inference speed reaching 44FPS. The results show that the model outperforms the existing methods in terms of accuracy, efficiency and robustness, which is of positive significance for computer-aided diagnosis.