Sensitivity of neural networks to corruption of image classification

Abstract

Artificial intelligence (AI) systems are extensively used today in many fields. In the field of medicine, AI-systems are especially used for the segmentation and classification of medical images. As reliance on such AI-systems increases, it is important to verify that these systems are dependable and not sensitive to bias or other types of errors that may severely affect users and patients. This work investigates the sensitivity of the performance of AI-systems to labeling errors. Such investigation is performed by simulating intentional mislabeling of training images according to different values of a new parameter called "mislabeling balance" and a "corruption" parameter, and then measuring the accuracy of the AI-systems for every value of these parameters. The issues investigated in this work include the amount (percentage) of errors from which a substantial adverse effect on the performance of the AI-systems can be observed, and how unreliable labeling can be done in the training stage. The goals of this work are to raise ethical concerns regarding the various types of errors that can possibly find their way into AI-systems, to demonstrate the effect of training errors, and to encourage development of techniques that can cope with the problem of errors, especially for AI-systems that perform sensitive medical-related tasks.