Dr. AI: Adapting CNN Classification Training for the Technical and Social Challenges of Medical Diagnosis

Abstract

Diagnosing medical images is a time, cost, and labor intensive task traditionally only undertaken by an expert few. Fortunately, through the development of artificial intelligence (AI) and accessibility to large medical datasets, convolutional neural networks have become increasingly suitable for learning to conduct computer-assisted diagnosis (CAD). However, learning for medical classification comes with the unique challenges of low data volume, class imbalance, inconsistent labeling, and having fine image details differentiate multiple diagnoses. In this thesis, we design a training methodology specifically tailored to the medical domain by integrating transfer-learning, dataset cleaning, and synthetic data augmentation techniques. Through evaluation of color channel variations in images used to pre-train a model, implementation of an iterative dataset cleaning scheme, and use of DeepInversion to synthesize additional training data, small but compounding improvements to classification performance are shown. Finally, through the gained experience of developing a CAD methodology and contextualization of medical AI research in prevalent social and legal discussions, a set of privacy and bias conscious design principles are introduced