【摘要】 针对CT图像肺结节检测和分割模型复杂且精度低的问题，提出一种端到端的融合注意力特征的多任务肺结节检测和分割算法。利用多任务模型对肺结节检测和分割进行建模，实现模型参数的共享和计算复杂度的降低；提出残差注意力特征融合模块融合尺度和语义不一致的特征，获取更加丰富的特征信息；采用自适应多任务损失函数，实现主任务和辅助任务损失权重的均衡。在LIDC-IDRI数据集上进行了详尽的实验，肺结节检测的CPM得分达到90.94%，肺结节分割的IoU和DSC分数分别为71.78%和80.89%，验证了算法的有效性。更多还原

【Abstract】 To address the problems of complex model and low accuracy of lung nodule detection and segmentation methods in CT images,an end-to-end multi-task lung nodule detection and segmentation algorithm fused with attention features was proposed.Multi-task learning model was used to handle the two tasks of lung nodule detection and segmentation,to realize the sharing of model parameters and the reduction of computational complexity.Residual attention feature fusion module was designed to fuse features with inconsistent scale and semantics to obtain richer feature information.Adaptive multi-task loss function was provided to achieve the balance of the loss weight between the main task and the auxiliary task for nodule detection and segmentation.Results of detailed experiments on the LIDC-IDRI dataset show that the CPM score of 90.94%for lung nodule detection,and the IoU and DSC scores of 71.78%and 80.89%for lung nodule segmentation are obtained,respectively,verifying the effectiveness of the algorithm.更多还原