Echocardiographic Image Segmentation with Vision Transformers: A Comparative Analysis of Different Loss Functions

Deep learning (DL) algorithms have demonstrated their effectiveness in performing several automatic segmentation tasks, reaching high levels of accuracy. In particular, Vision Transformers (ViT) have recently emerged as a competing alternative to Convolutional Neural Networks (CNNs), which for many years remained state-of-the-art methods for segmentation. In numerous applications, ViTs demonstrated superior performance than CNNs, thanks to their capability to capture long-range relationships within images. Recent studies focusing on ViTs aimed to address some of the main drawbacks associated with these networks, i.e. the high computational cost and the substantial volume of data required for training and deployment. MedFormer stands out as one of the most recent ViT-based networks designed for this specific purpose. It has already been demonstrated capable of outperforming other ViTs and CNNs in segmentation tasks involving Magnetic Resonance (MRI) and Computer Tomography (CT) images. This architecture is highly scalable with data size and doesn't require any pretraining. Recently, we also demonstrated its effectiveness in applications involving ultrasound images. We trained and tested it on the CAMUS echocardiographic dataset. MedFormer reached, and in some cases surpassed, the performance of CNNs, even in scenarios where other ViT-based architectures fell short. However, ViTs, and MedFormer in particular, still remain relatively underexplored in the ultrasound imaging domain. To facilitate further studies involving this type of architectures, we propose a comparison of MedFormer performance in segmenting CAMUS images using different training strategies. One of the many factors that may influence model learning is the choice of a proper loss function. In the current study, we trained MedFormer with some of most common loss functions used in semantic segmentation and we compared the results in terms of Dice score.