Locally advanced rectal cancer (LARC) presents a formidable clinical challenge, with early recurrence (ER) often signalling a more aggressive tumour phenotype and poor prognosis. Traditional methods for risk assessment, based largely on clinical markers such as carcinoembryonic antigen (CEA) and MRI-assessed features, offer limited accuracy due to tumour heterogeneity. With up to 70% of rectal cancer patients diagnosed at an advanced stage, there is an urgent need for more precise and reliable predictive models. A recent study explored and compared the performance of clinical, radiomics, deep learning (DL) and fusion models in predicting ER in LARC patients, using multiparametric MRI data from a multicentre cohort.
Radiomics and Deep Learning for Improved Prognostic Accuracy
Radiomics and deep learning have emerged as powerful tools in oncological imaging. Radiomics involves the extraction of high-dimensional quantitative features from medical images, capturing spatial heterogeneity that may be invisible to the naked eye. Deep learning complements this by identifying complex, high-order patterns without predefined rules, often outperforming traditional feature engineering.
Must Read: MRI Radiomics for Lymphovascular Invasion Prediction in Rectal Cancer
In this study, radiomic and DL features were extracted from four MRI sequences: T2-weighted imaging (T2WI), diffusion-weighted imaging (DWI), T1-weighted imaging (T1WI) and contrast-enhanced T1-weighted imaging (CET1WI). After rigorous feature selection involving reproducibility analysis and dimensionality reduction techniques, eight radiomic and four DL features were retained for model development. Notably, features from CET1WI dominated both sets, underlining its significance in capturing recurrence-relevant information. A wavelet-transformed radiomic feature derived from CET1WI was particularly influential in predicting ER, reflecting the sequence’s ability to characterise tumour heterogeneity.
Fusion Modelling and Comparative Performance
To evaluate predictive performance, five models were constructed: a clinical model, a radiomics model, a DL model and two fusion models—one based on early feature-level fusion and the other on late decision-level fusion. All models were developed using the XGBoost algorithm, chosen for its robustness in handling high-dimensional data and preventing overfitting.
The late fusion model, which combined the output probabilities of the clinical, radiomics and DL models, consistently outperformed the others. It achieved area under the curve (AUC) values ranging from 0.863 to 0.880 across all cohorts and demonstrated superior calibration and clinical utility. By integrating predictions at the decision level, the late fusion model preserved the strengths of each modality while avoiding the potential redundancy and noise inherent in early fusion approaches. In contrast, although the early fusion model showed moderate improvement over single-modality models, it suffered from a mismatch between sensitivity and specificity, limiting its generalisability.
Kaplan-Meier survival analysis further validated the prognostic relevance of the late fusion model. Patients categorised as high risk based on its output had significantly worse recurrence-free survival (RFS), confirming its potential as a tool for preoperative stratification and postoperative management.
Clinical Implications and Future Directions
The results highlight the potential of multiparametric MRI-based fusion models to enhance the preoperative identification of patients at high risk of ER. By combining clinical biomarkers with imaging-derived radiomic and DL features, the late fusion model offers a non-invasive and scalable method for improving personalised treatment strategies. High-risk patients could benefit from intensified follow-up, tailored adjuvant therapies or inclusion in clinical trials for novel interventions.
Despite these promising findings, several limitations must be addressed. The retrospective design and relatively small sample size may limit generalisability. Future research should focus on validating these results in larger, prospective studies and exploring the integration of additional imaging modalities such as CT and PET to further enhance model accuracy. Moreover, the study did not examine recurrence site-specific predictions, which could inform more targeted surveillance strategies.
The integration of clinical, radiomic and deep learning features using a late fusion strategy offers a robust, non-invasive approach for predicting early recurrence in locally advanced rectal cancer. With superior performance over traditional models, this method could support risk stratification and personalised care planning, marking a significant advancement in the management of rectal cancer.
Source: European Journal of Radiology
Image Credit: iStock
