Locally advanced rectal cancer (LARC) continues to present challenges in treatment due to the high variability in patient outcomes following standard protocols. While total mesorectal excision combined with neoadjuvant chemoradiotherapy (nCRT) has improved local control, recurrence or metastasis remains common. Identifying patients at high risk of disease recurrence is essential to tailor treatment and improve long-term disease-free survival (DFS).
Multimodal MRI, particularly diffusion kurtosis imaging (DKI), has shown promise in capturing relevant tumour characteristics. However, existing approaches require manual processing, limiting scalability. A multitask deep learning network, MultiRecNet, was developed to automate DFS prediction using clinical data and MRI, potentially enabling better risk stratification and personalised care.
A Multitask Deep Learning Framework
MultiRecNet was designed to simultaneously perform tumour segmentation, recurrence classification and survival prediction in patients with LARC treated with nCRT. The network used multimodal MRI inputs—T2-weighted images and DKI-derived parameters (ADC, Dapp and Kapp)—combined with pretreatment and postoperative clinical indicators. The study involved 445 patients from three centres in China, with separate cohorts for training, internal testing and external validation. Multiple model variations were tested by adjusting the combinations of clinical and imaging data, with model performance evaluated through segmentation accuracy (Dice similarity coefficient), classification capability (area under the curve, AUC) and survival prediction (concordance index, C-index).
Must Read: AI Fusion Models Predict Early Recurrence in Rectal Cancer
The best-performing configuration integrated all imaging and clinical inputs, achieving a Dice coefficient of 0.72, a classification AUC of 0.97 and a C-index of 0.92 for DFS prediction in the internal testing set. External testing confirmed its robustness, with a C-index of 0.81. These results outperformed both traditional convolutional networks and other multitask models, including Transformer-based and Deep Multi-Task Survival networks.
Predictive Accuracy and Risk Stratification
Model comparisons revealed that the addition of DKI imaging significantly enhanced recurrence prediction. The unimodal T2-weighted model underperformed relative to the T2+DKI model, which had a notably higher AUC for recurrence classification. Incorporating pretreatment clinical data further improved performance, and the most comprehensive model—including postoperative pathology—achieved the highest metrics across tasks. In Kaplan-Meier analyses, the model successfully stratified patients into high- and low-risk groups with significant differences in DFS.
Subgroup analyses assessed the generalisability of MultiRecNet, showing consistent performance across patients with different tumour characteristics and treatment responses. Propensity score matching confirmed the model’s predictive power, even in populations with divergent clinical profiles. Importantly, the model was able to identify patients at risk of early recurrence regardless of initial treatment response, supporting its utility in both early treatment planning and follow-up decisions.
Clinical Implications and Future Directions
MultiRecNet addresses several limitations of previous radiomic and deep learning models. By unifying segmentation, classification and survival tasks in a single network, it eliminates the need for manual tumour labelling and mitigates the risk of overlooking peritumoural features. The multitask framework enhances learning by allowing shared attention to tumour-relevant regions across tasks. The integration of multimodal data ensures adaptability to different stages of clinical care, making the tool potentially valuable for both preoperative and postoperative decision-making.
The study demonstrated that recurrence prediction was more accurate when including postoperative pathology. Nonetheless, even baseline imaging and clinical data alone yielded high predictive accuracy, allowing for early risk stratification. This is particularly relevant for selecting patients who might benefit from non-surgical management strategies or intensified monitoring. Moreover, the model's use of probabilistic outputs supports nuanced survival predictions that account for time-to-event and censored data.
The development and validation of MultiRecNet mark a significant advance in predictive oncology for LARC. Through the integration of multimodal MRI and comprehensive clinical data, this multitask deep learning model delivers robust predictions for recurrence and survival. Its fully automated design improves scalability and clinical usability, offering a valuable tool for personalising treatment strategies. While prospective validation and broader application remain necessary, MultiRecNet has the potential to support precision medicine by optimising care pathways based on individual risk profiles.
Source: Radiology: Imaging Cancer
Image Credit: iStock
