Source-linked AI summary
Sam-Correction: Fully Adaptive Label Noise Reduction for Medical Image Segmentation
Takuro Shimaya, Masahiro Saiko
TL;DR
Noisy medical segmentation labels are difficult to curate, and existing correction can require additional clean masks. SAM-C uses SAM with prompts from a concurrently trained segmentation network to adaptively correct labels. Experiments report precise noise reduction and strong segmentation performance, with a stated boundary when annotations deviate too far from image texture.
Problem
Noisy medical segmentation masks are common, while SC requires additional well-curated masks that may be inaccessible and uses a single bias parameter.
Source: §1 · p. 1Method
SAM-C iteratively uses SAM and prompts from a concurrently trained segmentation network to correct noisy labels without additional clean masks.
Source: §1 · p. 1Results
SAM-C outperformed existing methods in trained-network prediction accuracy and clean-GT reproducibility, and its mid-correction played a critical role.
Source: §4 · p. 4Takeaways & Limitations
SAM-C achieved precise noise reduction and high-performing segmentation without additional well-curated masks, including settings where multiple clinical experts are unavailable.
Source: §1 · p. 1For LIDC-IDRI “worst” masks, SC produced better corrected-mask DSC; SAM-C may struggle when labels deviate too far from image texture.
Source: §4 · p. 4
Abstract
from arXiv · showhide
The quality of the teacher label has a dramatic impact on the outcome of the machine learning process. Medical image segmentation often suffers from a noisy ground truth (GT) mask due to difficulties in consistent annotation. Although a recent study has proposed a promising method to curate the noisy GT mask, this requires additional well-curated GT masks that are not always accessible without the aid of multiple medical experts. As an alternative, we propose SAM-Correction, which utilizes a foundation model to correct the noisy GT mask in a fully adaptive manner instead of manually preparing well-curated masks. In our approach, we progressively mitigate label noise via the segment anything model (SAM), where prompts are given from another segmentation network that we train concurrently. Experiments with various types of label noise derived from actual human annotations demonstrate that SAM-C achieves precise noise reduction and high-performing segmentation models trained thereby. Our method paves the way for robust model training even when clean labels are entirely unavailable.
1. INTRODUCTION
Medical image segmentation labels are often noisy, while pixelwise approaches do not account for spatial noise and segmentation-specific methods remain limited. SAM-C uses SAM with prompts from a concurrently trained network to adaptively correct labels without additional clean masks; experiments report improved segmentation and noise reduction.
1. INTRODUCTION
- Spatial correlation limits pixelwise classification approaches to segmentation, where dedicated methods have made relatively little progress.Medical annotation noise can arise from ambiguous boundaries, required expertise, or limited annotation time.Source: §1 · p. 1
- SC estimates a single bias parameter using manually curated clean masks, which can be costly to obtain and may not capture nonuniform noise.Such masks are often based on consensus among several experts and are not always accessible.Source: §1 · p. 1
- SAM-C iteratively uses SAM and prompts from a concurrently trained segmentation network to correct noisy labels without extra clean masks; it outperformed SC and reduced noise precisely.The reported experiments use two datasets with non-synthetic noise.Source: §1 · p. 1
2. METHODS
SAM-C first uses SAM to correct noisy masks, trains a segmentation network on those labels, and then uses its predictions to prompt further correction. The study evaluates this process on two datasets with human annotations, comparing it with baselines and correction ablations using DSC.
2.1. SAM-C
- Pre-correction prompts SAM with boxes and masks from noisy-mask objects; mid-correction uses network predictions as prompts, filters nonoverlapping boxes with B-mix, and can repeat for Nc cycles.The network is trained on pre-corrected masks before mid-correction and retrained using its corrected masks.Source: §2.1 · pp. 1–2
2.2. Datasets
- Experiments use non-synthetic annotations from LIDC-IDRI and Cityscapes, with three LIDC-IDRI noise types and coarse Cityscapes annotations.LIDC-IDRI test and validation clean masks use majority voting; Cityscapes coarse annotations lack detailed margin structure.Source: §2.2 · p. 2
2.3. Experiments
- The study compares SAM-C with a noisy-mask baseline, two SC validation settings, and an oracle across four noise conditions, evaluating prediction and corrected-mask DSC.Ablations omit pre- or mid-correction and also test post-correction and combinations; LIDC-IDRI uses one cycle with B-mix, while Cityscapes uses two cycles without it.Source: §2.3 · p. 3, Table 1 · p. 3
3. RESULTS
SAM-C achieved the highest DSC among the proposed methods and performed closest to the oracle. Its correction processes improved segmentation and more robustly reduced annotation noise than SC, while post-correction was unstable across conditions.
3. RESULTS
- SAM-C achieved the highest DSC among proposed methods and came closest to the oracle, with Fig. 2 illustrating predictions under noisy annotations.The oracle achieved the highest accuracy overall.Source: §3 · p. 3, Figure 2 · p. 3
- SC with single TV improved LIDC-IDRI accuracy but not consistently Cityscapes, suggesting too few curated masks may limit robust performance.This variability may reflect overfitting the bias parameter and affects label-acquisition cost.Source: §3 · p. 3
- Pre-, mid-, and post-correction each helped individually, but pre- plus mid-correction robustly outperformed SC while post-correction combinations were unstable.Mid-correction outperformed pre-correction; post-correction sometimes achieved the highest DSC but was sometimes inferior to SC.Source: §3 · pp. 3–4
- SAM-C robustly reduced annotation noise in corrected teacher masks, whereas SC sometimes produced masks with lower DSC than the original noisy labels.SAM uses image texture for adaptive correction, unlike SC’s typical-bias-length expansion or shrinkage.Source: §3 · p. 4, Table 2 · p. 4
4. DISCUSSION AND CONCLUSION
SAM-C outperformed existing methods in prediction accuracy and clean-mask reproducibility without additional well-curated masks. Its limitation is performance on annotations that depart too far from image texture.
4. DISCUSSION AND CONCLUSION
- Without additional well-curated masks, SAM-C outperformed existing methods in DNN prediction accuracy and reproducibility of clean training masks.The concurrently trained DNN’s prompts made mid-correction a critical component.Source: §4 · p. 4
- For the worst LIDC-IDRI noisy masks, SC produced corrected masks with better DSC than SAM-C when annotations deviated too far from image texture.The authors suggest stronger or domain-compatible foundation models as possible remedies.Source: §4 · p. 4
5. COMPLIANCE WITH ETHICAL STANDARDS
Ethical approval was not required for the open data used.
5. COMPLIANCE WITH ETHICAL STANDARDS
- Ethical approval was not required because the open data’s license confirmed it was unnecessary.The stated basis was the license attached to the data.Source: §5 · p. 4