Pixels to Predictions: Designing Computational Imaging AI to Enable Precision Medicine

About this Class

Developing artificial intelligence (AI) schemes to assist the clinician towards enabling precision medicine approaches requires development of objective markers that are predictive of disease response to treatment or prognostic of longer-term patient survival.

The solutions being developed involve designing computational imaging features together with histology or molecular data for detailed tissue and disease characterization in vivo as well associated with patient outcomes. The key innovation in this approach lies in “handcrafting” unique tools that can capture biologically relevant and clinically intuitive measurements from routinely acquired imaging (MRI, CT, PET) or digitized images of tissue specimens.

Further, by conducting cross-scale associations between imaging, pathology, and -omics, we can not only “unlock” and integrate the information captured by these different, disparate data modalities, but also develop an interpretable and intuitive understanding of what drives their performance.

Specific problems addressed via the new computerized imaging markers we have developed through our NCI funded RadxTools project include:

(a) Predicting response to treatment to identify optimal therapeutic pathways

(b) Evaluating therapeutic response to guide follow-up procedures. Further examining how to account for differences between sites, scanners, and acquisition parameters to ensure generalizable performance of AI tools and computational imaging features, crucial for wider clinical translation and widespread adoption.

These will be discussed in the context of clinical applications in colorectal and renal cancers as well as digestive diseases.