INTRODUCTION:

Personalized management of locally advanced rectal cancer (LARC) depends on reliable prognostic prediction, which remains a challenge for current methods. Recently, novel deep learning models demonstrate exciting results in this field. This study was to develop and validate a multi-task deep learning model for full process automatic prediction of disease-free survival (DFS) of LARC patients treated with neoadjuvant chemoradiotherapy (NCRT).

METHODS

This study retrospectively included rectal cancer patients treated with nCRT from three centers, and each patient has clinical information, baseline multimodal MRI images (T2, ADC, Dapp, Kapp), and follow-up survival data. We developed a deep multi-task joint survival model that simultaneously performed tumor segmentation, risk classification, and survival prediction. These three tasks collaborated with each other, which contributed to fully exploiting the key features of both imaging data and clinical data. To better explore the multi-view information from multimodal data, an attention mechanism was introduced to learn the jointly represented features within and across different modalities, leading to reduced noise level on the experimental results. To cope with the missing data situation, we employed a mask-based graph network structure, which can fully utilize the data and improve the adaptability and robustness of the network. In addition, we compared the performance of other well-designed single-task models and our proposed model. Several metrics were adopted for evaluation, including DSC (Dice Similarity Coefficient) for segmentation task; AUC (area under the receiver operating characteristic curve), accuracy, sensitivity, specificity, positive predictive value and negative predictive value for risk classification task; C-index and AUC of time-ROC for survival prediction task.

RESULTS

There were 261, 75 and 73 cases included in the training, internal testing and external testing set respectively. The best performance for DFS prediction was presented on model with all available data input, achieving AUC of 0.961 (95%CI: 0.877-0.988) for risk classification, C-index of 0.915 (95% CI: 0.860, 0.970) for survival prediction, and DSC of 0.724 for segmentation. The high-risk group stratified by risk according to the DL-score of models had significantly worse DFS than the low-risk group (HR = 41.240 [95% CI: 13.777, 123.447], P < 0.001). And models’ generalizability was proven in the external testing set (HR = 4.195 [95%CI: 1.732-10.158, P < 0.001]).

DISCUSSION

In our study, the multi-task DL model performed significantly better than any single-task CNNs with manually labelling in relapse risk classification. We speculated that the inclusion of segmentation could lead to a more focused attention for both the classification and survival tasks. Although the area of attention did not exactly match the area labeled by the clinicians, it did not significantly deviate from the tumor area and may be beneficial in detecting valuable features of the peri-tumor area that were not recognized by the clinicians, which contributed to the improvement on predictive performance. In fact, several previous studies have demonstrated a tight association between image features in the peri-tumor region and tumor biology or clinical outcomes.

CONCLUSION

We developed and validated a fully automatic and multimodal deep learning model to predict tumor recurrence of LARC after nCRT without any manually labeling of the tumor. The model could integrate multiple types of prognosis-related information and effectively stratify the recurrence risk of patients at different clinical treatment periods, thus providing important information for treatment decisions-making and advancing the development of precision medicine in rectal cancer.

Acknowledgements

No acknowledgement found.

References

No reference found.