Advancing Veterinary Cytology with Deep Learning: Development, Validation, and Best Practices

Abstract

Recent technological advances have led to rapid growth in both digital and computational pathology. With this growth, the field of veterinary medicine has seen a significant expansion of diagnostics in computational pathology, specifically in artificial intelligence. Yet, there is little information on test validation, verification, or how quality assurance should be maintained for these new diagnostics. Historically, in veterinary medicine, test validation and verification and quality assurance guidelines have been set forth by the American Society of Veterinary Clinical Pathology (ASVCP) to ensure that high-quality diagnostic tests remain the standard. These guidelines do not address the unique needs of quality assurance in computational pathology. In recognition of this need, several consensus statements have been published targeting best practices for verification, validation, and quality assurance of these new diagnostics in the medical field. In the present work, we review the current literature to introduce pathologists to basic concepts of artificial intelligence-based diagnostics. We also introduce the minimal requirements for verification, validation, and quality assurance of artificial intelligence systems in the medical field. By providing high-quality diagnostics and standardizing quality assurance, we maintain trust and reliability in the diagnostic tests we, as a profession, offer. This allows us to better serve our patients, clients, and community while advancing veterinary medicine in a way that benefits all. Despite rapid advancements in computational pathology, little is known about the ideal magnification, image type, and number of images needed to train an artificial intelligence system, particularly when using images from cytology. Therefore, this manuscript also explores optimum conditions for building an artificial intelligence system called a convolutional neural network. The findings of this research suggest that higher magnification with either color or greyscale images is ideal for building a convolutional neural network using cytology samples. Additionally, the ideal number of images to use for a two-class problem is 150 images per class. This work is foundational in understanding the requirements of convolutional neural networks and allows for future studies.