Improved Multitask Model based on TransUNet in the Neoadjuvant Therapy for Rectal Cancer
10.11783/j.issn.1002-3674.2025.01.001
- VernacularTitle:基于TransUNet改进的多任务模型在直肠癌新辅助治疗中的应用
- Author:
Shuwen YIN
1
;
Zhipeng DING
;
Yan LI
Author Information
1. 哈尔滨医科大学公共卫生学院卫生统计学教研室(150081)
- Publication Type:Journal Article
- Keywords:
Deep learning;
Multitask learning;
Colorectal cancer;
Neoadjuvant therapy
- From:
Chinese Journal of Health Statistics
2025;42(1):2-6
- CountryChina
- Language:Chinese
-
Abstract:
Objective We improved the segmentation model TransUNet based on deep learning methods to construct a multitask model that can both segment regions of interest and predict classification,which made it possible to identify sensitive populations of patients undergoing neoadjuvant therapy for rectal cancer.Methods The multitask model added classification structure on the basis of TransUNet,including the fully connected layer(input of 512,output of 256),the ReLU activation function,and the fully connected layer(input of 256,output of 3),to achieve the prediction of triple classification outcomes(stable disease(SD),partial response(PD),and complete response(CR)).The 3D MRI of 71 rectal cancer patients before neoadjuvant chemotherapy admitted to the Harbin Medical University Cancer Hospital from 2015 to 2017 were extracted into 2D images as data for the study.Dice coefficient and Hausdorff distance were used to evaluate thesegmentation performance,and accuracy,micro-precision,micro-recall,and micro-F1 score were used for classification performance.Results The model was trained for 100 epochs,and the average Dice coefficient and average Hausdorff distance for the segmentation task on the test set were 0.851 and 10.806,respectively.For the classification task,the accuracy,micro-precision,micro-recall,and micro-F1 score on the test set were all 0.651 from the slicing perspective,and all four metrics were 0.857 from the patient perspective.Conclusion Our model works well in the segmentation task.Although the model performs poorly on the classification task at the slice level,the performance was acceptable at the patient level taking into account the tripartite classification results.The multitask model has the potential to be used in the clinic for assisted diagnosis.
