Source-linked AI summary
Semi-supervised few-shot learning for medical image segmentation
Abdur R Feyjie, Reza Azad, Marco Pedersoli, Claude Kauffman, Ismail Ben Ayed, Jose Dolz
TL;DR
Medical few-shot segmentation needs effective learning from very few pixel-labeled images without overfitting. The paper adds unlabeled surrogate tasks to episodic training, and reports stronger representations and improved skin-lesion segmentation on ISIC and PH2, especially in the 1-shot setting.
Problem
Medical segmentation models require many pixel-level masks, while low-data training increases overfitting risk; few-shot learning addresses this by learning novel classes from few labeled examples.
Source: §1.1 · p. 3Method
The framework combines episodic few-shot segmentation with an unlabeled denoising surrogate task that shares the encoder for semantic feature learning.
Source: §1.1 · p. 4, §1.2 · p. 4, §3 · p. 8Results
The surrogate-task approach improves few-shot segmentation on ISIC and PH2 by 6–7% over the few-shot baseline in the 1-shot setting and 3–4% in the 5-shot setting.
Source: §2.3 · pp. 6–7Takeaways & Limitations
Unlabeled surrogate tasks produce richer, more generic representations that improve generalizability to unseen segmentation tasks.
Source: §2.3 · p. 7, §3 · p. 8
Abstract
from arXiv · showhide
Recent years have witnessed the great progress of deep neural networks on semantic segmentation, particularly in medical imaging. Nevertheless, training high-performing models require large amounts of pixel-level ground truth masks, which can be prohibitive to obtain in the medical domain. Furthermore, training such models in a low-data regime highly increases the risk of overfitting. Recent attempts to alleviate the need for large annotated datasets have developed training strategies under the few-shot learning paradigm, which addresses this shortcoming by learning a novel class from only a few labeled examples. In this context, a segmentation model is trained on episodes, which represent different segmentation problems, each of them trained with a very small labeled dataset. In this work, we propose a novel few-shot learning framework for semantic segmentation, where unlabeled images are also made available at each episode. To handle this new learning paradigm, we propose to include surrogate tasks that can leverage very powerful supervisory signals --derived from the data itself-- for semantic feature learning. We show that including unlabeled surrogate tasks in the episodic training leads to more powerful feature representations, which ultimately results in better generability to unseen tasks. We demonstrate the efficiency of our method in the task of skin lesion segmentation in two publicly available datasets. Furthermore, our approach is general and model-agnostic, which can be combined with different deep architectures.
1 Methodology
The framework combines episodic few-shot segmentation with an unlabeled surrogate task, using shared representations to improve transfer to novel medical-image classes.
1.1 Few shot semantic segmentation
- Few-shot segmentation trains on disjoint base and novel classes, using k labeled support references and query images within episodic tasks.The model encodes support and query images, then predicts the query mask for the target class.Source: §1.1 · p. 3
- Support and query features are encoded by a CNN, and mask average pooling converts known support foreground regions into class prototypes for query prediction.The prototypes are unpooled to the query-feature resolution and convolved with query features before optimizing the segmentation loss.Source: §1.1 · p. 4
1.2 Surrogate task
- The surrogate task addresses few-shot feature-learning difficulty by adding an auxiliary objective whose network shares the encoder with the segmentation model.The surrogate-task parameters are separate from the few-shot decoder parameters.Source: §1.2 · p. 4, §1.3 · p. 5
- Image denoising corrupts unlabeled images with random noise and trains an encoder-decoder to reconstruct the originals using cross-entropy.Denoising is selected because rotation prediction is less suitable for skin tumors, whose orientations can vary widely; cross-entropy was empirically preferred over L2 reconstruction.Source: §1.2 · pp. 4–5
1.3 Joint objective
- The joint objective minimizes few-shot segmentation and surrogate-task losses, with λ controlling the surrogate task’s contribution.The segmentation loss depends on encoder and decoder parameters, while the surrogate loss depends on encoder and surrogate-network parameters.Source: §1.3 · p. 5
2 Experiments
Experiments evaluate the proposed few-shot segmentation framework on ISIC and PH2 using episodic training, unlabeled surrogate tasks, and qualitative and quantitative comparisons. The results indicate that surrogate-task supervision improves segmentation, especially with fewer labeled examples.
2.1 Dataset
- The experiments use FSS-1000 for base training and ISIC and PH2 as medical evaluation datasets, with unlabeled images supporting auxiliary tasks.ISIC includes an independent 1000-image evaluation set, while PH2 contains 200 dermoscopic images used only for testing.Source: §2.1 · pp. 5–6
2.2 Experimental Set-up
- The model uses a VGG-based encoder with atrous convolutions, a two-block decoder, DSC evaluation, and end-to-end Adam training.Images are resized to 224 × 224 pixels, and training runs for 30K iterations over 500 episodes.Source: §2.2 · p. 6
2.3 Results
- 6-7% improvement over the few-shot baseline was observed in the 1-shot setting across ISIC and PH2 when unlabeled surrogate tasks were used.Episodic training also improved performance by nearly 6% over regular batch-wise training.Source: §2.3 · p. 6, Table 1 · p. 7
- The results suggest that denoising-based surrogate tasks improve feature representations and yield larger gains when only one labeled example is available.Performance improvements are reported as more significant for 1-shot than 5-shot learning.Source: §2.3 · p. 7
- 3-4% improvement over few-shot training without the surrogate task was obtained in the 5-shot setting with 10 additional episode images.This gain was smaller than the 6-7% improvement reported for 1-shot learning.Source: §2.3 · p. 7, Table 2 · p. 7
- Qualitative results show progressively better segmentations from batch-wise to episodic training and then to episodic training with surrogate tasks.The proposed model increases true positives while reducing false positives and over-segmentation.Source: §2.3 · pp. 7–8, Figure 2 · p. 8
3 Conclusion
The conclusion presents episodic training and self-supervised surrogate tasks as a framework for few-shot medical image segmentation. Experiments on two public skin cancer datasets report improved performance and support transfer to unseen tasks.
3 Conclusion
- The work integrates episodic training into few-shot semantic segmentation for medical images and uses denoising surrogate tasks to learn transferable features.The surrogate task provides trivial supervision from unlabeled images while encouraging richer and more generic encoder representations.Source: §3 · p. 8
- Experiments on two public skin cancer segmentation datasets show significant improvements when unlabeled images are exploited through self-supervised tasks.The conclusion links enriched representations to better generalization on unseen tasks.Source: §3 · p. 8