Prediction of tumor regression grading in rectal cancer neoadjuvant chemoradiotherapy: a habitat radiomics analysis of imaging biomarker.
Authors
Affiliations (4)
Affiliations (4)
- Department of Radiation Oncology, Shandong Cancer Hospital and Institute, Shandong First Medical University and Shandong Academy of Medical Sciences, No. 440 Jiyan Road, Huaiyin District, Jinan, 250117, China.
- Department of Radiation Medicine, Shandong First Medical University and Shandong Academy of Medical Sciences, Taian, China.
- School of Computer Science and Engineering, Southeast University, No. 2 Sipailou Road, Xuanwu District, Nanjing, 210096, China. [email protected].
- Department of Radiation Oncology, Shandong Cancer Hospital and Institute, Shandong First Medical University and Shandong Academy of Medical Sciences, No. 440 Jiyan Road, Huaiyin District, Jinan, 250117, China. [email protected].
Abstract
Tumor regression grading (TRG) is a core prognostic predictor of treatment outcomes in rectal cancer. Conventional TRG assessment methods are limited in capturing the full complexity of intratumoral heterogeneity. Advances in medical imaging, particularly radiomics and habitat-based analysis, hold promise the improve TRG prediction by quantitatively characterizing subregional tumor features. This study aimed to evaluate the performance of habitat radiomics in preoperatively predicting TRG in rectal cancer patients receiving neoadjuvant chemoradiotherapy (nCRT). Computed tomography (CT) images were analyzed to compare the predictive performance of conventional radiomics features and habitat-based analysis. Tumor regions of interest (ROIs) were segmented, extracting local imaging features. Voxel-level clustering was employed to identify distinct intratumoral subregions. Machine learning algorithms, including ExtraTrees, support vector machine (SVM), and Random Forest, were applied to predict TRG. For the conventional radiomics model, the ExtraTrees algorithm yielded superior performance, with AUCs of 0.912 and 0.817 in training and testing cohorts, respectively, outperforming SVM and Random Forest. The habitat model outperformed conventional radiomics model, while the combined model integrating habitat features and clinical variables yielded the optimal efficacy (training AUC = 0.916, test AUC = 0.833). In the binary classification task of TRG0 (pathologic complete response, pCR) vs. TRG1-2, the Habitat model achieved a test AUC of 0.884, and the combined model further reached 0.929. SHAP analysis identified that features from the H1 subregion and wavelet-transformed features were the top predictive contributors. Habitat-based radiomics, especially when integrated with clinical data, significantly improves the preoperative prediction of TRG in rectal cancer patients undergoing nCRT, providing a powerful tool to advance personalized oncology. Further validation in large-scale, multicenter, independent cohorts is warranted to facilitate the clinical translation of this approach.