融合多尺度特征和子空间注意力的黏膜下肿瘤检测

计算机辅助检测工具可以帮助医生减少在临床检查中漏检误检的情况，从而提高诊断准确度，同时减轻医生的劳动强度。针对超声胃肠镜检查中黏膜下肿瘤的定位与分类问题，提出了一种融合多尺度特征和子空间注意力的黏膜下肿瘤检测算法(MFSA-YOLOv7t)。首先，移除小目标预测头，在保证精度下使网络轻量化；然后，基于浅层特征提出多尺度特征融合模块，提取肿瘤细节信息；其次，改进上采样结构，在保留上层信息的同时增强感受野；最后，引入子空间位置注意力模块，捕获肿瘤的位置和边界特征，进一步提升黏膜下肿瘤的检测性能。实验表明，MFSA-YOLOv7t在平均精度均值、敏感度以及准确度上分别达到97.32%，96.99%和96.24%，相比YOLOv7-tiny算法检测性能有较大的提升，分别提高了2.39%，2.75%和2.59%。MFSA-YOLOv7t为医生在临床检查中的辅助诊断提供更加可靠的肿瘤类型参考，同时为黏膜下肿瘤的检测提供了一个新的思路和研究方向。

Detection method in submucosal tumor based on multi-scale feature and subspace attention fusion

Computer-aided detection tools can help doctors reduce the situations of missed and false detections in clinical examinations to improve diagnostic accuracy and reduce the labor intensity of doctors. Aiming at the problem of the localization and identification of submucosal tumors in ultrasound gastroscopy, a submucosal tumor detection algorithm combined with multi-scale feature and subspace attention (MFSA-YOLOv7t) was proposed. Firstly, the small target prediction head was removed to make the network lightweight while maintaining accuracy. Secondly, A multi-scale feature fusion module was proposed based on shallow features to aggregate details. Then, the algorithm improved the up-sampling structure to retain the upper layer information to the greatest extent and enhance the feature perception field. Finally, the coordinate subspace attention module was introduced to capture tumor location and boundary features to improve the detection performance of submucosal tumors. Experiments show that the mAP, sensitivity, and accuracy of MFSA-YOLOv7t reach 97.32%, 96.99%, and 96.24%, respectively, compared with YOLOv7-tiny, MFSA-YOLOv7t has a great improvement in detection performance, which is improved by 2.39%, 2.75%, and 2.59%, respectively. MFSA-YOLOv7t provides a more reliable reference of tumor type for doctors in clinical examination of the auxiliary diagnosis and also provides a new idea and a research direction for detecting submucosal tumors.