PROJECT SUMMARY
Cutaneous squamous cell carcinoma (cSCC) is a highly-prevalent form of skin cancer with an estimated 1.8M
new diagnoses annually in the United States alone. While typically less aggressive than many other forms of
cancer, this high prevalence still results in significant levels of mortality, including approximately 15,000 deaths
each year in the U.S. This mortality risk combined with the large volume of cSCC cases creates a tremendous
need for optimizing treatment escalation strategies, so that patients at risk for poor outcomes can benefit from
more aggressive treatments while low-risk patients can be spared the corresponding side effects and costs.
Here, we propose to develop artificial intelligence (AI) algorithms to predict cSCC prognosis and ultimately aid
in treatment selection. The models will be trained and validated using a unique dataset of 1,360 cSCC tumors
with corresponding histopathology slides, clinical outcomes, and labels for known risk factors. The
histopathology whole slide images will be used as model input, where we propose a curriculum learning
strategy to mitigate overfitting while training in a weakly-supervised, multiple instance learning fashion. We will
first develop models to predict the presence of known risk factors, including poor differentiation, desmoplasia,
and the invasion of certain histologic structures. While these factors have been shown to correlate with poor
prognosis, they are challenging to consistently detect in clinical practice. Next, we will develop AI models to
directly predict prognosis, exploring the potential of these models to identify prognostic biomarkers not
previously identified. We will subsequently validate the performance of these models and use interpretability
techniques to investigate the features learned. This validation includes a clinical integration simulation study,
where the risk predictions of the AI models will be compared to the retrospective use of adjuvant therapy in
clinical practice. Altogether, accomplishment of this proposal would serve as initial steps in reducing the burden
of a disease that affects millions of patients per year, while developing methods to address common challenges
in computational pathology and especially the important task of prognosis prediction.
