Comparison of Loss functions and Optimizers for Multi-class X-ray Bone Segmentation

Abstract

X-ray bone segmentation helps orthopaedic surgeons make proper decisions by separating bones from soft tissues and making the view clear. Segmenting the bones help them to analyze if the bones are in place. UNet architectures are widely used for segmentation tasks. Selecting optimal configuration help in better segmentation of bones. This paper compared different optimizers and loss functions while studying pelvic and femur bone segmentation from X-ray images. Overall, AdamW optimizers yield better performance with different loss functions than all other optimizers, including the commonly used Adam. Tversky loss shows good stable results across different optimizers in terms of the loss function. Best dice similarity coefficient and intersection over union score of 97.04 % and 96.56 % are achieved using AdamW and dice loss.