Submission
| Title: |
Predicting Melanoma Metastasis with Simple Machine Learning Models that Use Human Interpretable Image Features Extracted from Histopathological Images |
| Co-Authors: |
Couetil, Justin, Indiana University School of Medicine; Ziyu Liu, Indiana University School of Medicine; Jie Zhang, Indiana University School of Medicine; Kun Huang, Ahmed Alomari, Indiana University School of Medicine |
Abstract
Background/Significance/Rationale: Melanoma is the fifth most common cancer overall, with ~106,110 new cases in 2021 in the US alone, and incidence rising about 1.4% each year (NCI, 2021). Though melanoma is the deadliest skin cancer, most people are cured when their superficial, thin, primary melanoma is surgically removed, with a 99.4% 5-year survival (NCI, 2021). These are the Stage I patients, and they make up 80% of all melanoma patients. The issue is, melanoma can come back years later (in the same region of the body or as distant metastasis), and currently there is no reliable method to predict whether or when a recurrence will happen.
Methods: We use interpretable machine learning models (i.e. logarithmic classification) and human-interpretable histopathology image features (e.g., nuclear area, staining intensity, lymphocyte density) to predict melanoma survival and metastasis risk.
Results/Findings: With our methods, we can predict metastasis within 5 years with 82% accuracy. With the public cohorts of late-stage patients, we predicted survival beyond 2.2 years with 81% accuracy.
Conclusions/Discussion: We also provide a visualization tool to highlight tissue regions which have many of the cell morphologies associated with good and bad prognoses. This enables pathologists to directly associate cell morphology and tissue architecture with the progression of cancer. This framework recapitulated the known poor-prognostic feature of tumor ulceration. In our future work, we plan to investigate the association between high-risk cell morphologies with bulk/spatial gene expressions, and to focus our attention towards gathering a large group of Stage I patients, so that our training data better reflects the real-world prevalence of melanoma stages.
Translational/Human Health Impact: Being able to predict metastasis in Stage I patients would be a paradigm shift in the management of melanoma – potentially catching locoregional and distant metastasis recurrence at an early stage, where curative surgical resection is possible.
