Nat Commun: Multi-task deep learning method to predict the treatment response of patients with solid tumors

Accurately predicting the treatment response of individual patients is essential for the patient's personalized drug treatment strategy.
In view of the non-invasive nature of radiography, it usually measures the changes in tumor size before and after treatment to evaluate the treatment response of solid tumors.

Due to the complexity and heterogeneity of treatment response patterns, this simple imaging method cannot always accurately assess potential biological responses.
Therefore, it is urgent to explore reliable methods for predicting tumor response.

In this study, the researchers proposed a multi-task deep learning method that can simultaneously predict tumor segmentation and response.
The researchers designed two Siamese sub-networks connected in multiple layers to integrate multi-scale representations and make in-depth comparisons between pre-processed and post-processed images.

Research process and research design

The network trained 2568 MRI scans of 321 rectal cancer patients to predict the pathological complete response after neoadjuvant chemoradiation.
The results showed that the imaging-based model achieved AUC (area under the ROC curve) of 0.
95 and 0.
92 in two independent cohorts of 160 and 141 patients, respectively.
When combined with blood-based tumor markers, the comprehensive model showed an AUC value of 0.
97 and further improved the accuracy of prediction.

Therapeutic response prediction network model

All in all, the results of the study reveal that the method of capturing dynamic information in longitudinal images can be widely used for solid tumor patient screening , treatment response assessment, and disease monitoring.

Original source:

Jin, C.