Source-linked AI summary
Prompt-Guided Interactive Segmentation of Interstitial Lung Disease in Thoracic CT
Vasilis Dedousis, Lubnaa Abdur Rahman, Lorenzo Brigatο, Ethan Dack, Andreas Christe, Christoph Frank, Manuela Funke-Chambour, Justus Roos, Adrian Huber, Lukas Ebner, Stavroula Mougiakakou
TL;DR
ILD segmentation remains difficult to refine interactively because existing methods depend on dense annotations and static predictions, while promptable-model adaptation to ILD is limited. This paper adapts MedSAM2 for interactive 3D thoracic-CT segmentation, evaluates fine-tuning and clinical prompt types, and tests automatic-mask initialization. Full fine-tuning was strongest, with a 4.7-percentage-point average Dice improvement over MedSAM2; BBox prompts led, while lasso and scribble remained effective, and the automatic-prior workflow was demonstrated as a proof of concept.
Problem
ILD segmentation methods remain limited by dense annotation requirements and static predictions, while promptable-model adaptation to ILD remains largely unexplored.
Method
The paper adapts MedSAM2 for interactive 3D ILD segmentation, evaluating three fine-tuning strategies and point, BBox, lasso, and scribble prompts.
Results
4.7 percentage points: full model fine-tuning improved average Dice over MedSAM2, while BBox prompts performed best and lasso and scribble remained effective.
Takeaways & Limitations
An automatic segmentation prior can initialize MedSAM2 for subsequent radiologist-prompted refinement in a proof-of-concept end-to-end workflow.
Takeaways & Limitations
Prospective clinical validation, external-dataset robustness, and improved propagation for highly fragmented ILD patterns remain necessary future evaluations.
Abstract
from arXiv · showhide
Accurate segmentation of interstitial lung disease (ILD) patterns is essential for quantitative disease assessment and longitudinal monitoring. However, existing approaches remain limited by relying on dense annotations and producing static predictions that cannot be refined, motivating interactive approaches. While promptable models show promise in interactive segmentation, their adaptation to ILDs remains largely unexplored. To address this gap, we investigate prompt-guided foundation models for ILD refinement and present, to the best of our knowledge, the first adaptation of MedSAM2 for interactive 3D ILD segmentation on thoracic CT. We investigate three fine-tuning strategies and multiple clinically motivated prompts: bounding-boxes (BBox), point, lasso, and scribble. On a dataset spanning seven ILD patterns and healthy lung tissue, full model fine-tuning performed best, improving the average Dice score by 4.7 percentage points over MedSAM2.While BBox prompts achieve the strongest performance, non-native MedSAM2 interactions such as lasso and scribble prompts also prove effective. Finally, we present and evaluate a proof-of-concept end-to-end workflow in which MedSAM2 is initialized from an automatic segmentation prior and subsequently refined using radiologist prompts. Model weights and plug-ins made available at: https://github.com/AIHNlab/ILD-SemiSegTool.
1 Introduction
ILD segmentation is difficult because abnormalities are fragmented, variably distributed, and poorly bounded, while conventional methods rely on static predictions. The paper therefore adapts MedSAM2 for prompt-guided interactive refinement on thoracic CT.
- Motivation: ILD patterns such as GGO, HC, and consolidation have fragmented distributions, poorly defined boundaries, and substantial appearance variability.Unlike lung lobes, ILD abnormalities lack fixed anatomical locations and boundaries.
- Motivation: These properties make automated ILD segmentation more challenging than segmentation of spatially coherent anatomical structures.
- Prompt-guided segmentation: Promptable models support human-in-the-loop segmentation through interactions such as points, bounding boxes, and masks.SAM2 additionally uses temporal memory to propagate masks across frames, a mechanism suited to volumetric CT.
- Contributions: The paper presents the first adaptation of MedSAM2 for interactive ILD segmentation on thoracic CT.
- Contributions: It evaluates point, BBox, scribble, and lasso interactions for refining diffuse pulmonary abnormalities.
- Contributions: A proof-of-concept workflow uses an automatic segmentation to initialize subsequent radiologist-guided MedSAM2 refinement.
2 Methods
The methods combine a multi-pattern ILD CT dataset with automatic nnU-Net initialization and prompt-guided MedSAM2 refinement. Experiments compare prompt types, fine-tuning strategies, propagation settings, and prompt polarity.
- 2.1 Dataset: The dataset contains 306 CT cases covering seven ILD patterns and healthy tissue, with ground truth annotated by two chest radiologists and resolved by consensus.
- 2.2 Segmentation: nnU-Net generates automated volumetric multi-class ILD segmentations that serve as initial mask priors for interactive refinement.
- 2.2 Segmentation: MedSAM2 initializes segmentation from prompts on a reference slice and propagates predictions bidirectionally through the volume.
- 2.2 Segmentation: The study evaluates point, single-BBox, multi-BBox, scribble, and lasso prompts, approximating scribble and lasso as point series because MedSAM2 was pretrained with BBox prompts.
- 2.2 Segmentation: Mask-prior refinement uses positive prompts for under-segmented regions and negative prompts for over-segmented regions, with P-only, N-only, and combined P+N settings evaluated.
- 2.2 Segmentation: Three fine-tuning strategies adapt the mask decoder only, mask decoder plus memory modules, or the full model.
- 2.3 Evaluation: Performance is assessed with Dice similarity coefficient and normalized surface distance using a 2 mm tolerance.
3 Results and Discussion
Full fine-tuning produced the strongest standalone MedSAM2 adaptation, while prompt effectiveness depended on ILD pattern and interaction setting. Mask-prior refinement improved average performance but showed complementary, not uniformly superior, behavior relative to automatic segmentation, with important dataset and workflow limitations.
- MedSAM2 adaptation: 0.380 DSC was the highest standalone score achieved by full fine-tuning, which also attained the highest NSD across prompt types.The NSD gains indicate sharper boundary delineation, and full fine-tuning was selected for subsequent experiments.
- Prompt strategy analysis: MBBox was the strongest prompt overall, while lasso and scribble performed comparably to native point prompts despite being absent from MedSAM2 pretraining.The full fine-tuned model maintained stable and statistically superior performance across all five prompt types.
- Prompting strategies: 2D interactions generally achieved higher DSC than 3D interactions, indicating a performance cost associated with reducing prompts to one propagated initialization slice.MBBox remained the most robust prompt in both settings, whereas points showed the greatest 2D-to-3D degradation across several patterns.
- Prompting strategies: MBBox performed best for fragmented multifocal patterns, whereas point-based prompts matched or exceeded BBox variants for healthy tissue.This pattern dependence reflects the irregular distribution of abnormalities versus the large, continuous healthy-lung structure.
- Prompt ablation: 0.448 DSC was achieved by FF+MP with positive and negative prompts, significantly exceeding the 0.303 DSC of nnInteractive with nnU-Net mask prior.Negative-only prompting scored highest among point-based interactions, while positive-only prompting consistently underperformed.
- Per-pattern and prompt polarity analysis: The coupled pipeline improved fragmented, erroneously segmented patterns but traded performance on patterns that nnU-Net already segmented well.Negative prompts preferentially improved Emph., GGO, and HC, whereas positive prompts performed best for Ret. and Ret.+GGO; Bronc. results were under-represented.
4 Conclusion
The study establishes the feasibility of interactive foundation models for 3D ILD segmentation and supports human-in-the-loop refinement for quantitative thoracic imaging. It also reports the study’s funding, computational resources, software contribution, and absence of competing interests.
- Findings: Full model fine-tuning achieved the strongest performance, while 3D propagation retained reasonable performance with substantially lower interaction burden.The authors also note that 2D slice-level interactions outperformed volumetric 3D propagation.
- Findings: An automatic segmentation prior can initialize MedSAM2 for subsequent prompt-guided refinement in a proof-of-concept workflow.This workflow uses dense prompting before radiologist-guided refinement.
- Implications: The findings support interactive foundation models as tools for clinically meaningful human-in-the-loop refinement rather than replacements for automatic algorithms.The stated application is quantitative thoracic imaging.
- Acknowledgments: The work was partly supported by the Swiss National Science Foundation through the PRISM-fILD project and used the University of Bern’s UBELIX HPC cluster.The acknowledgment also credits development of a custom OHIF plugin integrating the annotation algorithms.
- Disclosure: The authors declared no competing interests.This statement appears in the disclosure section.