ABSTRACT
Gastric cancer is a major global disease burden and leading cause of cancer mortality worldwide.
Current treatment decision is made primarily on the basis of staging, which divides patients into
several prognostic groups. For patients with localized and locally advanced disease, curative-intent
surgery with chemotherapy is the standard treatment. However, survival outcomes vary widely, even
among patients with disease of the same stage. Certain patients with early-stage disease have a
sufficiently low risk of recurrence and may not benefit from, or could even be harmed by,
chemotherapy given the associated toxicity and side effects. Conversely, many patients with
aggressive tumors do not respond well to standard chemotherapy and still recur despite receiving
extensive but ineffective treatment. Therefore, current one-size-fits-all approach is suboptimal,
leading to over- and under-treatment in many patients. There is an unmet need for reliable prognostic
and predictive models to guide personalized treatment of gastric cancer. To address this unmet need,
we propose robust radiomics features of tumor morphology and spatial heterogeneity and establish
their prognostic value. In addition, we will incorporate pathobiological knowledge into the design of
deep learning models for predicting prognosis. Further, we will develop novel deep learning
architecture to analyze longitudinal images for predicting pathologic response to neoadjuvant therapy.
Finally, by leveraging the complementary value of imaging data, clinicopathologic variables and serial
serum markers, we will construct integrative models to further improve prediction. If successful, the
proposed models will be useful in two ways: (1), identify which patients with early gastric cancer may
safely forego chemotherapy and avoid toxicity; (2), select the most effective chemotherapy regimen
for a given patient. Further, the models can also identify patients with advanced disease who do not
respond to standard chemotherapy and may benefit from novel targeted therapy or immunotherapy.
The proposed computational imaging approaches are generally applicable for response monitoring
and disease surveillance in many solid tumor types. Finally, the AI-based imaging technology
developed here can bring benefit to underserved populations in minority groups and community
settings. Progress made in gastric cancer will not only improve outcomes for patients in the US but
also have global impact given its high incidence and mortality worldwide.