Source-linked AI summary
Scribble2Label: Scribble-Supervised Cell Segmentation via Self-Generating Pseudo-Labels with Consistency
Hyeonsoo Lee, Won-Ki Jeong
TL;DR
Full cell segmentation labels are costly and labor-intensive to produce. Scribble2Label uses a few scribbles with prediction averaging and pseudo-label filtering, and the authors report that it outperforms existing state-of-the-art methods across modalities and scribble-detail levels.
Problem
Deep-learning cell segmentation commonly relies on full ground-truth labels, which are time-consuming and labor-intensive to generate and can contain pixel-level errors.
Source: §1 · p. 2Method
Scribble2Label uses a few manual scribbles, averages predictions during training, and filters pseudo-labels to generate labels for unscribbled pixels.
Source: §1 · p. 2, §2.1 · p. 4, §2.2 · p. 5Results
Scribble2Label consistently outperforms existing state-of-the-art methods across various cell image modalities and levels of scribble detail.
Source: §1 · p. 2Takeaways & Limitations
The authors conclude that Scribble2Label can effectively avoid time-consuming, labor-intensive manual label generation.
Source: §4 · p. 9
Abstract
from arXiv · showhide
Segmentation is a fundamental process in microscopic cell image analysis. With the advent of recent advances in deep learning, more accurate and high-throughput cell segmentation has become feasible. However, most existing deep learning-based cell segmentation algorithms require fully annotated ground-truth cell labels, which are time-consuming and labor-intensive to generate. In this paper, we introduce Scribble2Label, a novel weakly-supervised cell segmentation framework that exploits only a handful of scribble annotations without full segmentation labels. The core idea is to combine pseudo-labeling and label filtering to generate reliable labels from weak supervision. For this, we leverage the consistency of predictions by iteratively averaging the predictions to improve pseudo labels. We demonstrate the performance of Scribble2Label by comparing it to several state-of-the-art cell segmentation methods with various cell image modalities, including bright-field, fluorescence, and electron microscopy. We also show that our method performs robustly across different levels of scribble details, which confirms that only a few scribble annotations are required in real-use cases.
1 Introduction
Scribble2Label uses a few scribbles to generate and filter training labels, addressing the cost of full pixel-level annotations and the lack of scribble-supervised methods for cell segmentation. The authors report performance over existing state-of-the-art methods across microscopy modalities and levels of scribble detail.
1 Introduction
- Full ground-truth labels are time-consuming and labor-intensive to generate, and manual pixel-level masks can be error-prone.Prior point-annotation methods also require annotations across the dataset and are sensitive to point placement; scribble-supervised methods had not yet been applied to cell segmentation.Source: §1 · p. 2
- The method combines pseudo-labeling with label filtering and prediction consistency to generate training labels from a small set of manual scribbles.The authors say this process removes noise in pseudo-labels and improves prediction accuracy.Source: §1 · p. 2
- Scribble2Label consistently outperforms existing state-of-the-art methods across cell image modalities and different levels of scribble detail.The introduction presents robustness across scribble detail as evidence that only a few manual annotations are needed.Source: §1 · p. 2
- Scribble2Label is an end-to-end scheme that needs no additional model parameters or external segmentation methods during training.The introduction identifies this as a difference from existing scribble-supervised segmentation methods.Source: §1 · p. 3
- The paper presents Scribble2Label as the first scribble-supervised segmentation method applied to cell segmentation across microscopy images.The authors describe it as a weakly supervised method that works with only a handful of manual annotations.Source: §1 · pp. 2–3
2 Method
Scribble2Label first trains on scribbled pixels, then refines predictions using scribble loss and filtered pseudo-labels for unlabeled pixels. It averages predictions during training to generate the pseudo-labels.
2 Method
- Training has two stages: a scribble-only warm-up followed by refinement with both scribbled and unscribbled-pixel losses.The second-stage unscribbled loss uses filtered pseudo-labels.Source: §2 · p. 3
2.1 Warm-Up Stage
- During warm-up, the network trains only on scribbled pixels, while periodically averaging predictions for use in the next stage.The ensemble step uses unaugmented images and occurs every γ epochs.Source: §2.1 · pp. 3–5
2.2 Learning with a Self-Generated Pseudo-Label
- After warm-up, consistently predicted unlabeled pixels are filtered, one-hot encoded, and used as pseudo-labels for the unscribbled-pixel loss.The total loss combines scribbled and unscribbled losses; EMA continues during this training stage.Source: §2.2 · p. 5
3 Experiments
Experiments evaluate Scribble2Label across cell-image datasets and compare it with fully supervised and weakly supervised methods. The reported results include strong performance across modalities and stability as scribble detail varies.
3.1 Datasets
- Evaluation covered MoNuSeg, three DSB modality subsets, and an internally collected larval-zebrafish EM dataset.The DSB subsets were bright-field, fluorescence, and histopathology; scribbles were manually drawn with reference to full labels, with EM scribbles generated at a 10% ratio.Source: §3.1 · pp. 5–6
3.2 Implementation Details
- Training used pretrained U-Net with ResNet-50, switching to ResNet-34 for histopathology comparisons, plus RAdam and conventional augmentation.For noisier MoNuSeg, the consistency threshold was 0.95 and EMA alpha was 0.1, rather than the general 0.8 and 0.2 settings.Source: §3.2 · p. 6
3.3 Results
- Scribble2Label outperformed all compared methods on IoU and mDice, and exceeded full-label mDice on EM, DSB-BF, and DSB-Histo.The MoNuSeg results were also described as outstanding for instance segmentation despite its small, cluttered cells.Source: §3.3 · p. 7, Table 1 · p. 7
- Prediction averaging was reported as crucial for robust pseudo-labels, which, combined with scribbles, supported accurate boundaries comparable to ground-truth labels.The comparisons also report noisy Grab-Cut and pseudo-label proposals, while rLoss sometimes failed with complex backgrounds.Source: §3.3 · pp. 7–8
- On DSB-Fluo, Scribble2Label produced stable semantic and instance results from 10% to 100% of skeleton pixels.At 30% of skeleton pixels, its performance was comparable to using full segmentation masks.Source: §3.3 · p. 8, Table 2 · p. 8
4 Conclusion
Scribble2Label combines pseudo-labeling, label filtering, and consistency for scribble-supervised cell segmentation. The conclusion reports accurate performance across datasets and scribble-detail levels without extra segmentation processes or model parameters.
4 Conclusion
- Scribble2Label combines pseudo-labeling and label filtering with consistency for scribble-supervised segmentation.The conclusion reports accurate performance across datasets and scribble-detail levels without extra segmentation processes or additional model parameters.Source: §4 · pp. 8–9
- The authors envision the method could help avoid time-consuming and labor-intensive manual label generation.Source: §4 · p. 9
- Future directions include extending the method beyond cell segmentation and generating labels for tumor regions and mitochondria.The paper names semantic and instance segmentation in images and videos as broader settings of interest.Source: §4 · p. 9