Support teaching, research, and patient care.
Dr. Novoa received his bachelor and medical degrees from Harvard University. He completed his medical internship at Columbia University Medical Center and his dermatology residency at University Hospitals-Case Western Reserve School of Medicine before pursuing a dermatopathology fellowship at the University of Pennsylvania. Board certified in dermatology and dermatopathology, Dr. Novoa practices clinical dermatology and interprets slides as a dermatopathologist. His research interests include the medical applications of artificial intelligence, cutaneous lymphoma, and the cutaneous side effects of targeted therapies. Dr. Novoa served as co-PI on a research project featured on the cover of Nature Magazine. This work has appeared in the Wall Street Journal, Bloomberg Magazine, and PBS NOVA. He enjoys travel, reading, and Brazilian Jiu Jitsu.
Recent advances in artificial intelligence stand poised to transform a number of human endeavors. A new approach, known as deep learning, harnesses large datasets, increased computing power, and convolutional neural networks to enable algorithms to discern complex patterns in raw data. These algorithms are widely used in autonomous vehicles, language translation, and image classification. Within medicine, these algorithms have been applied to a wide range of questions, ranging from detection of atrial fibrillation or breast cancer to prediction of ICU mortality. This same approach to pattern recognition can be applied to visual diagnosis of skin lesions, including melanoma.Currently, melanoma is responsible for nearly 10,000 annual deaths in the United States alone, and early diagnosis is critical to cure. Furthermore, melanoma mortality demonstrates various disparities, with worse outcomes for patients of low socioeconomic status or those living in rural areas. In addition, keratinocyte carcinomas are the most common human malignancy, with over 5 million cases a year in the United States and thousands of deaths from aggressive local or metastatic disease. Using deep learning techniques, we created a convolutional neural network (CNN) trained on 129,000 skin images and a morphological taxonomy composed of over 2000 disease categories. In a proof-of-concept study, our CNN demonstrated classification performance on par with 21 board-certified dermatologists across a range of tasks, including clinical and dermoscopic evaluation of melanocytic lesions. On expanded validation datasets, the algorithms showed comparable performance. More remains to be done, however, in order to validate this algorithm and examine its performance on prospective lesions. Furthermore, deep learning techniques struggle with 1) bias; 2) interpretability; 3) a lack of data; and 4) challenges with adversarial examples, where minute changes in angle or lighting can produce radically different outputs. In the next steps of our research, we seek to tackle these challenges while working on a prospective clinical trial of this technology in the real world.