Abstract
Objective:
To develop an effective machine learning model to predicts T-stage downstaging in MRI-cT4 stage rectal cancer (RC) patients before neoadjuvant therapy.
Materials and Methods:
Data from 136 MRI-T4 RC patients (96 males [70.6%]; mean age ± standard deviation, 54.7±12.3 years) were randomly divided into the training (n = 108) and test cohort (n = 28) with a ratio of 8:2. Based on the comparison between MRI-T grade before neoadjuvant therapy and pathological-T grade, patients were divided into Well T-downstage (cT4 to ypT≤2) and Poor T-downstage (cT4 to ypT≥3) groups. A total of 1834 radiomics features were extracted from T2-weighted imaging (T2WI) images. The least absolute shrinkage and selection operator (LASSO) regression method was applicated to select features. k-Nearest Neighbor (k-NN) was applicated to construct radiomics and clinical models. Finally, a model called RCMix that combined clinical with radiomics features was developed. The RCMix model was evaluated by receiver operating characteristic (ROC) curve and decision curve.
Results:
RCMix model was developed by CEA, CA19-9, and 14 radiomic features. The area under the ROC curve (AUC) of the RCMix model was 0.918 (95 % confidence interval, 0.868–0.967) in the training cohort and 0.836 (0.669–1.000) in the test cohort. The AUC of the radiomics model was 0.746 (0.656–0.837) in the training cohort and 0.813 (0.669–0.957) in the test cohort. The AUC of the clinical model was 0.847 (0.780–0.914) in the training cohort and 0.778 (0.593–0.962) in the test cohort. The AUC of the RCMix model was significantly larger than that of the clinical model (p = 0.049). The decision curve analysis indicated its clinical usefulness.
Conclusion:
The RCMix model achieved satisfactory performance in predicting T-stage downstaging in MRI-cT4 RC patients, accurately stratifying this group prior to receiving neoadjuvant therapy.

Feiyu Bai

Appearances