Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

ULS+: Data-driven Model Adaptation Enhances Lesion Segmentation

Rianne Weber, Niels Rocholl, Max de Grauw, Mathias Prokop, Ewoud Smit, Alessa Hering

TL;DR

ULS+ enhances interactive whole-body lesion segmentation by integrating six additional fully annotated public datasets and introducing click-point augmentation, paired with a reduced input size to accelerate inference. The method improves Dice similarity and robustness to click-point variations over the original ULS baseline, achieving first place on the ULS23 leaderboard and enabling faster, clinically tractable inference (~0.5 s on an A100 with TTA). Key innovations include data-driven expansion of training data, lesion-shifting augmentation, and test-time augmentation, all implemented within nnUNet v2 on $128 \times 128 \times 64$ inputs. The work demonstrates that continual, data-centered updates can yield robust, scalable improvements for longitudinal lesion quantification, with reproducible code available to support adoption and further development.

Abstract

In this study, we present ULS+, an enhanced version of the Universal Lesion Segmentation (ULS) model. The original ULS model segments lesions across the whole body in CT scans given volumes of interest (VOIs) centered around a click-point. Since its release, several new public datasets have become available that can further improve model performance. ULS+ incorporates these additional datasets and uses smaller input image sizes, resulting in higher accuracy and faster inference. We compared ULS and ULS+ using the Dice score and robustness to click-point location on the ULS23 Challenge test data and a subset of the Longitudinal-CT dataset. In all comparisons, ULS+ significantly outperformed ULS. Additionally, ULS+ ranks first on the ULS23 Challenge test-phase leaderboard. By maintaining a cycle of data-driven updates and clinical validation, ULS+ establishes a foundation for robust and clinically relevant lesion segmentation models.

ULS+: Data-driven Model Adaptation Enhances Lesion Segmentation

TL;DR

ULS+ enhances interactive whole-body lesion segmentation by integrating six additional fully annotated public datasets and introducing click-point augmentation, paired with a reduced input size to accelerate inference. The method improves Dice similarity and robustness to click-point variations over the original ULS baseline, achieving first place on the ULS23 leaderboard and enabling faster, clinically tractable inference (~0.5 s on an A100 with TTA). Key innovations include data-driven expansion of training data, lesion-shifting augmentation, and test-time augmentation, all implemented within nnUNet v2 on inputs. The work demonstrates that continual, data-centered updates can yield robust, scalable improvements for longitudinal lesion quantification, with reproducible code available to support adoption and further development.

Abstract

In this study, we present ULS+, an enhanced version of the Universal Lesion Segmentation (ULS) model. The original ULS model segments lesions across the whole body in CT scans given volumes of interest (VOIs) centered around a click-point. Since its release, several new public datasets have become available that can further improve model performance. ULS+ incorporates these additional datasets and uses smaller input image sizes, resulting in higher accuracy and faster inference. We compared ULS and ULS+ using the Dice score and robustness to click-point location on the ULS23 Challenge test data and a subset of the Longitudinal-CT dataset. In all comparisons, ULS+ significantly outperformed ULS. Additionally, ULS+ ranks first on the ULS23 Challenge test-phase leaderboard. By maintaining a cycle of data-driven updates and clinical validation, ULS+ establishes a foundation for robust and clinically relevant lesion segmentation models.
Paper Structure (8 sections, 1 equation, 4 figures, 2 tables)

This paper contains 8 sections, 1 equation, 4 figures, 2 tables.

Table of Contents

  1. Introduction
  2. Materials and methods
  3. Datasets
  4. Lesion shifting
  5. Training
  6. Evaluation
  7. Results
  8. Discussion

Figures (4)

  • Figure 1.1: Differences between the ULS model and the ULS+ model.
  • Figure 1.2: Dice scores of the original ULS model and ULS+ on two different test sets, stratified by lesion location.
  • Figure 1.3: Robustness scores (as defined in Eq. \ref{['3886-equation']}) of the original ULS model and ULS+ on two different test sets, stratified by lesion location.
  • Figure 1.4: Example segmentations of the ULS and ULS+ model on both datasets on lesions from different locations. Blue: ground truth, red: ULS, yellow: ULS+. Note that for visibility of this figure, some lesions are zoomed in on while preserving surrounding context. Therefore, in this visualization, some lesions may not appear centered.