Application of Microsatellite Instability in Endometrial Cancer via A Prediction Model Based on Diffusion Weighted Imaging Deep Learning Features
10.3969/j.issn.1005-5185.2024.09.011
- VernacularTitle:基于DWI深度学习特征的预测模型评估子宫内膜癌微卫星不稳定状态的价值
- Author:
Yongchao NIU
1
;
Fang ZHOU
;
Dandan ZHAO
;
Mengyan HOU
;
Shujian LI
;
Yong ZHANG
Author Information
1. 新乡市中心医院磁共振科,新乡市医学影像工程技术研究中心,河南新乡 453000
- Keywords:
Endometrial neoplasms;
Magnetic resonance imaging;
Diffusion weighted imaging;
Deep learning;
Radiomics;
Microsatellite instability
- From:
Chinese Journal of Medical Imaging
2024;32(9):922-927
- CountryChina
- Language:Chinese
-
Abstract:
Purpose To explore the value of a prediction model based on diffusion weighted imaging(DWI)deep learning features in endometrial cancer microsatellite instability status assessment.Materials and Methods DWI data of 32 microsatellite instability and 55 microsatellite stability endometrial cancer patients were analysed from June 2020 to April 2023 in Xinxiang Central Hospital,retrospectively.Apparent diffusion coefficient(ADC)values of the primary lesions were measured,and deep learning features and imaging histological features of the primary lesions were extracted using multilayer convolutional neural networks and PyRadiomics,respectively.The least absolute shrinkage and selection operator and random forest were used for feature screening and model building,respectively.The area under the receiver operating characteristic curve(AUC)and net reclassification improvement were used to evaluate model performance.Bootstrap based on 1 000 resamples was used for internal validation of the model.Results For the deep learning model,a total of 6 features were included,the 7th,57th,77th,82nd,97th and 108th features,with an AUC of 0.905(95％CI 0.823-0.957);for the radiomics model,a total of 6 features were included,1 neighborhood grey level difference matrix,4 grey level region size matrices and 1 grey level tour length matrix feature,with an AUC was 0.844(95％CI 0.751-0.913);for ADC values,the microsatellite instability group had smaller ADC values than the microsatellite stability group(t=-4.123,P＜0.001),with an AUC of 0.810(95％CI 0.712-0.886).Compared with the radiomics model and ADC values,the deep learning model showed improved risk prediction,with net reclassification improvements of 0.856 and 0.486(P＜0.01,P=0.024),respectively.In Bootstrap-based internal validation,the deep learning model also demonstrated higher performance than the radiomics model,with AUCs of 0.897(95％CI0.889-0.905)and 0.829(95％CI0.812-0.839),respectively.Conclusion A prediction model based on deep learning features of DWI images can provide a better assessment of microsatellite instability status in endometrial cancer patients than radiomics model and ADC values.
