Prognostic Prediction Value of the Clinical-Radiomics Tumour-Stroma Ratio in Locally Advanced Rectal Cancer
Chongpeng Cai,Tingdan Hu,Zening Rong,Jing Gong,Tong
DOI: https://doi.org/10.1016/j.ejrad.2023.111254
IF: 4.531
2024-01-01
European Journal of Radiology
Abstract:Purpose: To develop and validate a radiomics model based on high-resolution T2WI and a clinical-radiomics model for tumour-stroma ratio (TSR) evaluation with a gold standard of TSR evaluated by rectal specimens without therapeutic interference and further apply them in prognosis prediction of locally advanced rectal cancer (LARC) patients who received neoadjuvant chemoradiotherapy. Methods: A total of 178 patients (mean age: 59.35, range 20-85 years; 65 women and 113 men) with rectal cancer who received surgery alone from January 2016 to October 2020 were enrolled and randomly separated at a ratio of 7:3 into training and validation sets. A senior radiologist reviewed after 2 readers manually delineated the whole tumour in consensus on preoperative high-resolution T2WI in the training set. A total of 1046 features were then extracted, and recursive feature elimination embedded with leave-one-out cross validation was applied to select features, with which an MR-TSR evaluation model was built containing 6 filtered features via a support vector machine classifier trained by comparing patients' pathological TSR. Stepwise logistic regression was employed to integrate clinical factors with the radiomics model (Fusion-TSR) in the training set. Later, the MR-TSR and Fusion-TSR models were replicated in the validation set for diagnostic effectiveness evaluation. Subsequently, 243 patients (mean age: 53.74, range 23-74 years; 63 women and 180 men) with LARC from October 2012 to September 2017 who were treated with NCRT prior to surgery and underwent standard pre-treatment rectal MR examination were enrolled. The MR-TSR and Fusion-TSR were applied, and the Kaplan-Meier method and log-rank test were used to compare the survival of patients with different MR-TSR and Fusion-TSR. Cox proportional hazards regression was used to calculate the hazard ratio (HR). Results: Both the MR-TSR and Fusion-TSR models were validated with favourable diagnostic power: the AUC of the MR-TSR was 0.77 (p = 0.01; accuracy = 69.8 %, sensitivity = 88.9 %, specificity = 65.9 %, PPV = 34.8 %, NPV = 96.7 %), while the AUC of the Fusion-TSR was 0.76 (p = 0.014; accuracy = 67.9 %, sensitivity = 88.9 %, specificity = 63.6 %, PPV = 33.3 %, NPV = 96.6 %), outperforming their effectiveness in the training set: the AUC of the MR-TSR was 0.65 (p = 0.035; accuracy = 66.4 %, sensitivity = 61.9 %, specificity = 67.3 %, PPV = 27.7 %, NPV = 90.0 %), while the AUC of the Fusion-TSR was 0.73 (p = 0.001; accuracy = 73.6 %, sensitivity = 71.4 %, specificity = 74.0 %, PPV = 35.73 %, NPV = 92.8 %). With further prognostic analysis, the MR-TSR was validated as a significant prognostic factor for DFS in LARC patients treated with NCRT (p = 0.020, HR = 1.662, 95 % CI = 1.077-2.565), while the Fusion-TSR was a significant prognostic factor for OS (p = 0.005, HR = 2.373, 95 % CI = 1.281-4.396).Conclusions: We developed and validated a radiomics TSR and a clinical-radiomics TSR model and successfully applied them to better risk stratification for LARC patients receiving NCRT and for better decision making.