Resolving impact of technical and biological variability on the convolutional neural networks: evaluating chest x-ray scans

Abstract

Large chest x-ray datasets accumulate technical and biological variations. Technical variations are often caused by differences in scanner types (e.g. x-ray energy) whereas biological variations refer to those resulting from morphological differences among populations. Such variations decrease the generalizability of convolutional neural networks (CNNs) and could lead to disparities within specific patient subgroups. We sought to disentangle the sources of variations for a CNN in the context of technical and biological variability. The primary goal of this study was to understand whether tailored and specific AI models might need to be created (i.e., gender-specific or machine-specific AI models).

Ähnliche Arbeiten