Optimizing 3D UNet for real-time kidney segmentation in 3D ultrasound: a comparative study of loss functions
Simão Valente, Helena R. Torres, Pedro Morais, Andreas Fritz, Lukas R. Buschle, Estêvão Lima, João L. Vilaça · 2025
In the field of medical imaging, real-time segmentation of 3D ultrasound (US) volumes of the kidney is crucial for a variety of clinical applications. Recently, deep learning methods have been applied with excellent performance, where deeper and more complex networks are being proposed. However, increasing the complexity of the network can lead to challenges in maintaining real-time application. Therefore, loss functions play a very important role in optimizing state-of-the-art networks that not only demonstrate accurate segmentation performance but also remain suitable for real-time segmentation. This study investigates the impact of different loss functions on the training of 3DUNet for the segmentation of 3D US volumes of the kidney. The aim is to evaluate and compare ten loss functions to enhance the performance of the 3DUNet, while preserving its benefits of reduced computational cost and fast inference times. The research was conducted utilizing a dataset of 66 3D US volumes, randomly divided into 46 volumes for training, 10 for validations, and 10 for testing. Despite the different loss functions evaluated having received interesting segmentations results, the study demonstrated that a boundary-based loss was the most effective for this application. It achieves a superior balance of high Dice scores, low average surface distance (AverSSD) and competitive 95th percentile of the Hausdorff distance (HffSSD95), improving both segmentation quality and boundary precision, maintaining the computational resources of the 3DUNet. The overall findings of this study demonstrate that carefully selected loss functions can enhance network optimization, highlighting the feasibility of optimized deep learning networks for real-time kidney segmentation in 3D US imaging for clinical procedures.