Disentangled Latent Augmentation for Abnormality Detection in Musculoskeletal Radiographs

This chapter explores the application of disentangled latent augmentation for enhancing abnormality detection in musculoskeletal radiographs. The study addresses the challenge of class imbalance in medical datasets, which often skews learning towards the majority class and diminishes model sensitivity to rare but clinically important findings. The proposed methodology leverages a -Variational Autoencoder ( -VAE) architecture to learn structured latent representations from radiographs, enabling the generation of class-specific synthetic data. By incorporating triplet loss into the VAE framework, the approach promotes intra-class compactness and inter-class separation in the latent space. Experiments conducted on the MURA dataset, focusing on finger and forearm radiographs, demonstrate significant improvements in abnormality detection. The study highlights the effectiveness of the proposed method in generating high-quality, class-consistent synthetic samples that can be used to augment minority classes. The findings suggest a trade-off between interpretability and discriminative performance, with -VAE models offering unique advantages for generative and detection tasks. Future research directions include integrating synthetic abnormalities into semi-supervised and multi-modal pipelines, as well as enhancing -VAE with attention- or transformer-based encoders for improved representation quality and interpretability.