Development of a combined nomogram integrating radiomics, deep learning and clinical risk factors to predict neoadjuvant chemoradiotherapy outcomes in patients with rectal cancer

Technical Summary: Combined Nomogram for Predicting Neoadjuvant Chemoradiotherapy Outcomes in Rectal Cancer

Problem Statement
Accurate prediction of neoadjuvant chemoradiotherapy (nCRT) outcomes in locally advanced rectal cancer (RC) is critical for determining subsequent treatment strategies, such as "watch and wait" for complete responders or timely surgery for non-responders. Current evaluation methods, including pathological examination (invasive and limited by sampling), digital rectal exams (subjective and unable to quantify depth), and serum tumor markers like CEA (ineffective for non-secretors), lack the non-invasive, accurate, and personalized capabilities required for clinical decision-making. Furthermore, post-nCRT lesions often exhibit edema, fibrosis, and inflammation, complicating assessment via standard MRI alone. This study addresses the need for a robust, non-invasive predictive tool that integrates imaging and clinical data to forecast nCRT efficacy.

Methodology
The study utilized a retrospective cohort of 292 rectal cancer patients (230 in the training set, 62 in the validation set) who underwent pre-nCRT T2-weighted MRI and subsequent total mesorectal excision (TME). The outcome was defined by pathological tumor regression grading (TRG), categorizing patients into responders (TRG1/2) and non-responders (TRG3).

The methodology involved a multi-step modeling approach:

1. Image Preprocessing and Annotation: MR images underwent N4 bias field correction, resampling to isotropic 1mm³ voxels, and manual segmentation of intra-tumoral and peri-tumoral regions by experienced radiologists using ITK-SNAP.
2. Radiomics Model Development: 1,561 radiomics features (shape, first-order statistics, and texture) were extracted using PyRadiomics. Features were filtered for inter-observer reliability (ICC > 0.80), multicollinearity (|r| ≥ 0.95), and selected via LASSO regression. Two models were built: one using only intra-tumoral regions (Rad-Tumor) and one combining intra- and peri-tumoral regions (Rad-Tumor+Peri).
3. Deep Learning (DL) Model Development: A ResNet50 convolutional neural network was trained on the same segmented regions. Data augmentation (rotation, translation, scaling, flipping) was applied to increase the training sample size eightfold. Similar to radiomics, two DL models were constructed: DL-Tumor and DL-Tumor+Peri.
4. Clinical Variable Selection: Clinical factors (demographics, MRI staging, serum biomarkers) were screened via univariate and multivariate regression. Only Extramural Vascular Invasion (EMVI) and Tumor Location were retained as significant predictors.
5. Nomogram Construction: A combined nomogram was developed by integrating the Radscore (from Rad-Tumor+Peri), DLscore (from DL-Tumor+Peri), and the selected clinical variables (EMVI, Tumor Location).

Key Contributions

• Integration of Peri-tumoral Regions: The study demonstrates that including peri-tumoral adipose tissue features significantly enhances predictive performance over tumor-only models for both radiomics and deep learning approaches.
• Hybrid Modeling Strategy: The paper introduces a nomogram that synergizes high-dimensional deep learning features, statistical radiomics signatures, and specific clinical risk factors, rather than relying on a single modality.
• Comparative Analysis: It provides a direct performance comparison between traditional radiomics, deep learning, and their combined application within a single clinical framework.

Results

• Model Performance: In the validation dataset, the combined nomogram achieved the highest predictive accuracy with an Area Under the Curve (AUC) of 0.921 (95% CI, 0.824–0.974). This was significantly superior to the standalone DL-Tumor+Peri model (AUC 0.845) and the Rad-Tumor+Peri model (AUC 0.739) (all p-values < 0.05).
• Impact of Peri-tumoral Features: Incorporating peri-tumoral regions improved the AUC for both the radiomics model (from 0.666 to 0.739) and the deep learning model (from 0.817 to 0.845) in the validation set.
• Clinical Utility: Decision Curve Analysis (DCA) indicated that the nomogram provided a higher net benefit across the majority of threshold probabilities compared to other models. Calibration curves and the Hosmer–Lemeshow test confirmed good consistency between predicted probabilities and actual outcomes.
• Specific Clinical Factors: Multivariate analysis identified EMVI and Tumor Location as the only significant clinical predictors among the 13 factors initially collected.

Significance and Claims
The authors claim that the developed nomogram serves as an effective, non-invasive tool for predicting postoperative nCRT outcomes in rectal cancer patients. By integrating deep learning features, radiomics signatures, and clinical variables, the model offers improved diagnostic performance over single-modality approaches. The study posits that this tool can assist clinicians in personalizing treatment strategies, potentially sparing patients from unnecessary side effects of ineffective therapy or preventing tumor progression by identifying non-responders early. The authors acknowledge limitations, including the retrospective, single-center nature of the study and the complexity of the computational methods, suggesting that external validation is necessary to confirm robustness.