基于DCE-MRI影像的深度多任务学习联合预测乳腺癌病理信息研究  

Prediction of Breast Cancer Histological Information by Deep Multi-Task Learning Based on DCE-MRI Images

作　　者：袁成成 范明 许茂盛[2] 王世威[2] 厉力华[1] YUAN Chengcheng;FAN Ming;XU Maosheng;WANG Shiwei;LI Lihua(Institute of Biomedical Engineering and Instrument,Hangzhou Dianzi University,Hangzhou Zhejiang 310018,China;Department of Radiology,Zhejiang Hospital of Traditional Chinese Medicine,Hangzhou Zhejiang 310006,CAina)

机构地区：[1]杭州电子科技大学生物医学工程与仪器研究所,浙江杭州310018 [2]浙江省中医院放射科,浙江杭州310006

出　　处：《传感技术学报》2021年第3期354-360,共7页Chinese Journal of Sensors and Actuators

基　　金：国家自然科学基金项目(61871428,61731008);浙江省自然科学基金项目(J19H180004)。

摘　　要：乳腺癌病理信息中的Ki-67表达水平、分子分型及组织学分级等信息可为患者提供预后信息,本文提出一种基于深度多任务学习对乳腺癌多种病理信息联合预测的方法。回顾性分析202例术前,化疗前的浸润性导管癌患者乳腺DCE-MRI影像,将数据集随机分为训练集122例,测试集80例。分割影像的含肿瘤的腺体区域,并获得6个减影序列,利用深度学习结合多任务学习方法对Ki-67表达、分子分型及组织学分级中任意显著相关的两个任务进行联合预测,并计算AUC评估模型的分类性能。多任务学习影像模型预测Ki-67表达、Luminal A和组织学分级的最佳AUC为0.804,0.757,0.724。实验结果表明,深度多任务学习可以很好的预测乳腺癌病理信息的性能,对乳腺癌的诊断和个性化治疗方案的选择具有重要意义。The Ki-67 expression level,molecular typing and histological grade in breast cancer histological information can provide prognostic information for patients.This article proposes a method for prediction of histological information of breast cancer based on deep multi-task learning.We analyzed 202 cases of breast DCE-MRI images of patients with invasive ductal carcinoma before surgery and chemotherapy.The data set was randomly divided into 122 cases in training set and 80 cases in test set.Segment the tumor-containing gland area of the image,obtain 6 subtraction sequences,and use deep learning combined with multi-task learning methods to jointly predict any two tasks in Ki-67 expression,molecular typing and histological grading,the prediction performance was evaluated by AUC.The multi-task learning image model predicts the best AUC of Ki-67 expression,Luminal A and histological grades of 0.804,0.757,0.724.Experimental results show that deep multi-task learning can predict the performance of breast cancer histological information very well,which is of great significance for the diagnosis and selection of personalized treatment plan for breast cancer.

关 键 词：乳腺癌 深度学习 多任务学习 组织学分级 KI-67表达 分子分型 

分 类 号：TP393[自动化与计算机技术—计算机应用技术]