Deep learning prediction of chemo-immunotherapy response using tumor perfusion ultrasound images

Tumor heterogeneity poses a significant challenge for predicting responses to cancer therapy, highlighting the need for the development of biomarkers to guide personalized treatment. Contrast-enhanced ultrasound (CEUS) imaging is an established method to assess tumor perfusion, which directly affects drug delivery and therapeutic efficacy, as poorly perfused tumors often limit the penetration of chemo- and immunotherapeutics. Here, we developed a deep learning framework using CEUS imaging to predict tumor response to chemo-immunotherapy in murine models of breast cancer, fibrosarcoma, and melanoma. A convolutional neural network (CEUS-CNN) was trained on a dataset of 587 pre-treatment CEUS images to classify tumors as responsive, stable, or non-responsive based on RECIST criteria (175 responsive cases, 136 stable, and 276 non-responsive). Our model achieved an overall test accuracy of 0.859 (0.927 for responsive, 0.587 for stable, 0.948 for non-responsive) using only real data. Synthetic data were then generated for the responsive and stable classes to address class imbalance, leading to improved model performance, particularly for the previously underperforming stable class. While the non-responsive class maintained consistent accuracy, the responsive class experienced a decline. Finally, to enhance performance without compromising well-performing classes, synthetic augmentation was applied only to the underrepresented stable class. This targeted strategy enhanced model performance, raising the average test accuracy to 0.877 (0.943 for responsive, 0.686 for stable, 0.930 for non-responsive). These findings support CEUS imaging as a potential imaging biomarker of response to cancer therapy and highlight the promise of integrating AI with CEUS for personalized cancer treatment strategies.