Image processing approach provides robust feature extraction for classification with small sample sizes

Understanding the effects of chronic alcohol consumption on bone architecture is of great clinical importance due to its influence on skeletal health. Medical images contain valuable information for machine learning approaches to classify features relevant to alcohol use; however, the sample sizes are too small for traditional approaches. In this work, we develop a novel image feature extraction technique designed for small image datasets and apply it to analyze the effects of intrinsic (e.g., age, sex) and extrinsic (e.g., alcohol consumption patterns) factors on bone architecture. We train our models using images ascertained from microcomputed tomography on bone samples from humans and non-human primates. We achieve the best performance in both species when distinguishing bones from males and females (72% in macaque, 65.5% in human). We are able to distinguish between drinking and non-drinking individuals with an accuracy of 68% in macaques and 65% in humans, suggesting that our image processing approach is able to capture general biological features across species. Although the effects of alcohol on bone architecture are subtle, we find that they are detectable directly from imaging data.