Application of deep learning models based on super-resolution endorectal ultrasound in predicting perineural invasion in rectal cancer
10.3760/cma.j.cn131148-20250429-00244
- VernacularTitle:基于超分辨率直肠腔内超声的深度学习模型预测直肠癌周围神经浸润
- Author:
Yajiao GAN
1
;
Qiping HU
1
;
Xinyi WANG
1
;
Yixi SU
1
;
Qingling SHEN
1
;
Minling ZHUO
1
;
Yi TANG
1
;
Xiaodong LIN
1
;
Yue YU
1
;
Youjia LIN
1
;
Qingfu QIAN
1
;
Zhikui CHEN
1
Author Information
1. 福建医科大学附属协和医院超声科，福州 350001
- Publication Type:Journal Article
- Keywords:
Endorectal ultrasound;
Super-resolution reconstruction;
Rectal cancer;
Perineural invasion;
Deep learning
- From:
Chinese Journal of Ultrasonography
2025;34(10):848-857
- CountryChina
- Language:Chinese
-
Abstract:
Objective:To develop a deep learning model based on super-resolution endorectal ultrasound（ERUS）images for the preoperative prediction of perineural invasion（PNI）in patients with rectal cancer，thereby providing a reference for risk stratification and individualized treatment planning.Methods:A retrospective analysis was conducted on 382 patients with rectal cancer who underwent total mesorectal excision at Fujian Medical University Union Hospital between June 2019 and February 2024. Patients were randomly divided into a training set（ n=305）and a test set（ n=77）at a ratio of 8∶2，and further grouped into PNI-negative group and PNI-positive group subgroups based on pathological results. Super-resolution ultrasound images were generated from original ERUS images using a generative adversarial network（GAN）. Deep convolutional neural networks were developed based on features from intratumoral and peritumoral regions to identify the optimal region of interest（ROI）. The dSR5_ResNet18 and dSR5_ResNet50 models were constructed using the super-resolution images with a 5-pixel peritumoral extension. Representative clinical features were selected for subgroup analysis based on sample size and intergroup statistical differences between PNI-positive and PNI-negative patients. Forest plots were used to evaluate model applicability and robustness across subgroups. Results:The dSR5_ResNet18 model，built using super-resolution images of the tumor combined with a 5-pixel peritumoral region，achieved the best predictive performance，with an AUC of 0.867（95% CI=0.782 - 0.952）in the test set. Decision curve analysis demonstrated that the dSR5_ResNet18 model provided the greatest net clinical benefit. Forest plot analysis indicated strong generalizability of the models across subgroups such as pathological N stage，maximum lesion length，and lymph node enlargement，though relatively weaker performance was observed in the carcinoembryonic antigen（CEA）subgroup. Among all models，dSR5_ResNet18 exhibited the most consistent performance across subgroups，with the narrowest confidence intervals and highest robustness. Conclusions:The deep learning model incorporating ERUS-based super-resolution reconstruction demonstrated excellent performance in the preoperative prediction of PNI in rectal cancer. It offers significant advantages in image quality and generalizability，and may serve as a valuable tool to assist clinicians in formulating personalized treatment strategies.
