基于DPU-Net的直肠癌肿瘤分割与T分期研究

Rectal tumor segmentation and T staging based on DPU-Net

基于MRI图像对直肠癌肿瘤进行分割与T分期识别在直肠癌术前诊断与治疗方案制定中起着重要作用。为了准确分割肿瘤并进行T分期,提出一种多任务学习模型DPU-Net,联合分割与分期任务。在分割分支中,针对直肠癌MRI图像结构复杂的特点,结合注意力机制与多尺度特征加强模型对肿瘤的关注度以及边缘特征提取能力,提高分割效果。在分类分支中,为解决直肠癌肿瘤T分期难的问题,引入诊疗文本,充分利用医疗数据。提出基于动态加权的多模态融合模型,结合图像特征与文本特征对肿瘤T分期识别。将本文模型与主流模型进行对比,实验结果表明,本文模型分割精度DSC为82.88%,相比于U-Net提升了17.96%。分类准确率为76.24%,相比于Dense-Net提高了9.43%。本文模型具备辅助医生诊断的可行性。

The segmentation and T staging of rectal tumor in MRI images are critical in preoperative diagnosis and treatment planning.A multi-task learning model,DPU-Net,is proposed to achieve both accurate tumor segmentation and T staging.In the segmentation path,considering the complex structure of rectal cancer MRI images,attention mechanism and multi-scale features are used to enhance the model's focus on tumor and the edge feature extraction ability,thereby improving segmentation performance.In the classification path,medical treatment texts are introduced to make full use of medical data;and a multi-modality fusion model based on dynamic weight which combines image features and text features is established for T staging.The experimental results show that the proposed model achieves a Dice similarity coefficient of 82.88%,which is 17.96%higher than U-Net,and that the staging accuracy is 76.24%,which is 9.43%higher as compared with Dense-Net.The proposed method is feasible for auxiliary diagnosis.