Universal Lesion Segmentation Challenge 2023: A Comparative Research of Different Algorithms

TL;DR

The paper addresses universal lesion segmentation across multiple tissue types while prioritizing fast inference. It systematically compares nnUNetv2, DeepLabV3+, Medical Transformer, SwinUnet, and TransUNet on a 256×256×128 VOI, evaluating memory requirements, Dice-based accuracy, and cross-tissue robustness. Across the experiments, most architectures struggle with memory constraints and cross-tissue generalization, with TransUNet delivering the best performance after targeted fine-tuning though universal applicability remains elusive. The work highlights practical barriers to deployment and proposes future directions such as distributed training, model distillation, and knowledge-embedding approaches to approach truly universal segmentation.

Abstract

In recent years, machine learning algorithms have achieved much success in segmenting lesions across various tissues. There is, however, not one satisfying model that works well on all tissue types universally. In response to this need, we attempt to train a model that 1) works well on all tissue types, and 2) is capable of still performing fast inferences. To this end, we design our architectures, test multiple existing architectures, compare their results, and settle upon SwinUnet. We document our rationales, successes, and failures. Finally, we propose some further directions that we think are worth exploring. codes: https://github.com/KWFredShi/ULS2023NGKD.git

Figures (12)

• Figure 1: The Proposed Pipeline from ULS Challenge Host
• Figure 2: Baseline Result
• Figure 3: LOGO Learning and Gated Axial Attention
• Figure 4: SwinUNet Architecture
• Figure 5: SwinUNet Result