Deep learning integration of ultrasound perfusion and stiffness maps predicts tumor response to treatment

The complex tumor microenvironment and tumor heterogeneity make treatment responses highly variable, driving the shift toward biomarker-based therapy prediction. Ultrasound shear wave elastography (SWE) provides a simple and non-invasive method to measure tissue mechanical properties, such as tumor stiffness, by tracking shear wave propagation speed. Increased tumor stiffness is closely associated with compression of intratumoral blood vessels, which can severely limit blood flow within solid tumors, causing hypo-perfusion and hypoxia. These two abnormalities, in turn, hinder effective drug delivery and induce immunosuppression, which compromise the efficacy of cancer therapies. Tumor perfusion can be evaluated using contrast-enhanced ultrasound (CEUS), a minimally invasive imaging technique that employs microbubble contrast agents alongside ultrasound imaging to visualize blood flow and quantify tissue perfusion. In this study, we first developed CNN-ultra, a convolutional neural network trained separately on SWE and CEUS images to predict tumor response to chemo-immunotherapy in murine models of breast cancer, sarcoma, and melanoma. The dataset consisted of 587 SWE and CEUS images acquired prior to treatment, and tumor response was classified as responsive, stable, or non-responsive according to RECIST criteria (175 responsive, 136 stable, and 276 non-responsive cases). The overall accuracy of CNN-ultra was 0.855 for the SWE images and 0.876 for the CEUS images. Subsequently, we developed CNN-combi, an integrated model designed to combine features from both SWE and CEUS images to enhance predictive performance. Our CNN-combi model achieved an overall test accuracy of 0.912, with class-specific accuracies of 0.958 for responsive, 0.753 for stable, and 0.960 for non-responsive tumors, demonstrating the potential of the combination of SWE and CEUS as imaging biomarkers for predicting therapeutic response.