30 Stories in 30 Days
September is Thyroid Cancer Awareness Month. For the next 4 weeks, we will post stories written by thyroid cancer survivors, caregivers and friends for our 30 Stories in 30 Days campaign. We hope their perspectives and insight will help others along their journey.
On the cutting edge of the fields of pathology, oncology, and computer technology, researchers at Mount Sinai West and the University of Buffalo are using artificial intelligence to improve the classification of papillary thyroid carcinoma. Margaret Brandwein-Weber MD, Pathology Site Chair, Assistant Professor Scott Doyle PhD, a biomedical engineer in the Department of Pathology and Anatomical Sciences at the University of Buffalo, and their teams are teaching computers to recognize thyroid cancer from digitalized pathology slides. They are focusing on the tall cell variant of Papillary Thyroid Cancer, which describes cells that are at least twice as tall as they are wide. This variant is included in the American Thyroid Association Guidelines for Risk Stratification, as it is associated with increased risk of disease recurrence. Interestingly, this is the first time that this inherent geometry problem is being examined with a computational approach.
Currently, the degree of cancer “tallness” is visually estimated through microscopic study, which can be flawed due to individual subjectivity. Estimation can become particularly difficult when examining many slides from larger tumors, as the cancer’s appearance can differ from one region to another. Artificial intelligence could reduce this subjectivity and improve accuracy of classification.
The first aim in their project is the development of a machine classifier to calculate the range and distribution of tumor cell geometries. Their second aim is to look back at tumors defined as “Tall Cell Variant,” commonly defined as cancer composed of 50% or greater distribution of tumor cells which are at least twice as tall as they are wide. They will also study tumors classified as having “Tall Cell Features,” with thresholds below this 50% estimate. They will also look at patient outcomes from these two groups. Applying a machine classifier to digitalized tumor slides from patients with known outcomes offers a tremendous advantage and flexibility that is not possible with visual estimates. This approach allows for the examination of the spectrum of possible cutoffs for tumor cell geometry and distribution, as well as the determination of the best predictive cutoff values in diagnosing Tall Cell Variant and its associated clinical outcomes. The optimized machine classifier could then be developed into a digital companion prognostic test to classify patients into risk groups with accuracy and reproducibility. The ultimate goal is to customize patient-specific treatment plans to improve patient outcomes and quality of life.