TL;DR
This study introduces a novel "merged model" that combines clinical factors, MRI-based radiomics, and deep learning to predict pathological complete response (pCR) in locally advanced rectal cancer (LARC) patients undergoing neoadjuvant chemoradiotherapy (nCRT). The pre-treatment merged model demonstrated superior predictive performance and clinical utility compared to post-treatment models, offering a promising non-invasive tool for personalized treatment decisions.
What Problem Does This Paper Address?
Accurately predicting pathological complete response (pCR) in patients with locally advanced rectal cancer (LARC) after neoadjuvant chemoradiotherapy (nCRT) is crucial for optimizing treatment strategies. Currently, pCR is only definitively confirmed through post-surgical histopathology. Achieving pCR allows some patients to opt for a "Watch & Wait" approach, preserving quality of life and avoiding surgical complications without compromising survival. However, the variable response to nCRT makes it challenging to identify these patients pre-emptively. Existing methods like MRI-based tumor regression grading (mrTRG) have limitations due to subjective interpretation. While radiomics and deep learning (DL) have shown promise individually, their combined potential, especially in a merged model for predicting pCR before or after nCRT, remains underexplored. This paper aims to bridge this gap by developing and comparing pre- and post-nCRT merged models.
Methods
The study employed a multi-center approach with two cohorts: a training and test set (cohort 1, n=197) and an external validation set (cohort 2, n=52).
- Clinical Factor Analysis: Univariable logistic regression was used to identify significant clinical predictors of pCR.
- Radiomics Analysis: High-resolution T2W MRI images (pre- and post-nCRT) were used. A radiologist manually segmented the region of interest (ROI). A vast number of radiomics features (2264) were extracted and preprocessed. Inter- and intra-observer reliability was assessed using ICCs. Features with good robustness were further analyzed using LASSO regression to select optimal parameters for building "Radscore" models (Radscore-1 for pre-nCRT, Radscore-2 for post-nCRT).
- Deep Learning (DL) Model Construction: A Residual Network (ResNet) was utilized for 3D ROI classification. DL models (DL model-1 for pre-nCRT, DL model-2 for post-nCRT) were trained using the segmented radiomics data.
- Merged Models: Multivariate logistic regression was employed to combine selected clinical factors, the respective Radscore, and the DL model's predicted probability to create two merged models: Merged Model 1 (pre-nCRT) and Merged Model 2 (post-nCRT).
- Model Evaluation: Predictive performance was assessed using receiver operating characteristic (ROC) curves (AUCs), sensitivity, specificity, accuracy, precision, and F1 scores. Decision curve analysis (DCA) was used to evaluate clinical utility.
Datasets
- Cohort 1 (Training and Test Sets): 197 patients with LARC from Changhai Hospital. This cohort was split into an 80% training set and a 20% test set.
- Cohort 2 (External Validation Set): 52 patients with LARC from Ruijin Hospital Luwan branch. This cohort served as an independent validation set.
The study did not introduce new benchmarks but utilized existing patient data for model development and validation.
Key Insights and Ideas
- Pre-treatment Advantage: The pre-nCRT merged model (Merged Model 1) consistently outperformed the post-nCRT merged model (Merged Model 2) in predicting pCR across both test and validation sets. This suggests that information available before treatment is more robust for predicting the ultimate response.
- Synergy of Modalities: Combining clinical factors, radiomics (Radscore), and deep learning into a single "merged model" significantly enhanced predictive performance compared to any single modality alone. This highlights the complementary nature of these data sources.
- Clinical Utility Matters: Decision curve analysis (DCA) demonstrated that the pre-nCRT merged model provided a greater net clinical benefit across a range of threshold probabilities, indicating its practical value in guiding treatment decisions.
- Radiomics and DL Complementarity: While both radiomics and DL showed promise, their integration within the merged model amplified their individual strengths, leading to more accurate predictions.
Key Results
- Merged Model 1 (Pre-nCRT):
- Test Set: AUC of 0.834 (95% CI: 0.737–0.931), sensitivity of 97.8%, accuracy of 88.2%.
- Validation Set: AUC of 0.746 (95% CI: 0.636-0.856), sensitivity of 96.7%, accuracy of 80.9%.
- Merged Model 2 (Post-nCRT):
- Test Set: AUC of 0.746 (95% CI: 0.636-0.856), accuracy of 81.5%.
- Validation Set: AUC of 0.737 (95% CI: 0.646–0.828), accuracy of 75.8%.
- Individual Models: Pre-nCRT Radscore-1 and DL model-1 showed AUCs of 0.713 and 0.753 in the test set, respectively. Post-nCRT Radscore-2 and DL model-2 had AUCs of 0.653 and 0.663. These individual models performed less effectively than the merged models.
- Clinical Factors: CEA, MR T stage, and MR N stage were identified as independent predictors of pCR.
Limitations and Future Work
- Manual Segmentation: The study relied on manual segmentation of ROIs for radiomics analysis. While considered more accurate, this is time-consuming and may not scale well for very large datasets. Future work could explore automated segmentation.
- Retrospective Design and Sample Size: The retrospective nature and relatively small sample size of the validation cohort introduce potential selection bias.
- Limited MR Sequences: Only high-resolution T2W sequences were used. Incorporating other MR sequences (e.g., contrast-enhanced MRI) could potentially improve performance.
- Computational Burden: The authors acknowledge the need for transformer-based models that can handle large datasets more efficiently.
- Future Directions: The authors suggest that large, multi-center studies are warranted to decrease data bias and improve generalization. Further investigation into transformer-based models and the inclusion of additional MR sequences are also promising avenues. The developed pre-nCRT merged model is proposed as a promising tool to enhance diagnostic confidence for radiologists and clinicians in managing LARC cases.
Summary updated