[Paper Review] Semi-supervised few-shot learning for medical image segmentation

Abdur Razzaq Fayjie, Reza Azad+4 more|arXiv (Cornell University)|Mar 18, 2020

Domain Adaptation and Few-Shot LearningComputer Science27 references61 citations

TL;DR

A few-shot semantic segmentation framework that leverages unlabeled images via surrogate tasks to learn more transferable features for medical image segmentation, demonstrated on skin lesion datasets.

ABSTRACT

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.

Motivation & Objective

• Motivate the need for accurate medical image segmentation with limited pixel-level annotations.
• Propose a few-shot segmentation framework that also utilizes unlabeled images per episode.
• Introduce surrogate tasks derived from the data to provide strong supervisory signals.
• Demonstrate improved generalizability of learned features to unseen segmentation tasks.

Proposed method

• Train segmentation models in episodes representing different segmentation problems with few labeled examples.
• Incorporate unlabeled images per episode to enrich supervision.
• Introduce surrogate tasks that exploit data-derived supervisory signals for feature learning.
• Show that surrogate tasks improve feature representations and generalizability.
• Maintain model-agnostic compatibility so the approach can be combined with different architectures.

Experimental results

Research questions

• RQ1 Can unlabeled images used within each episode enhance performance in few-shot medical image segmentation?
• RQ2 Do surrogate tasks grounded in the data improve the learned feature representations for unseen segmentation tasks?
• RQ3 Is the proposed framework compatible with various deep segmentation architectures?
• RQ4 How does the method perform on public skin lesion segmentation datasets?

Key findings

• Including unlabeled surrogate tasks leads to more powerful feature representations.
• The method improves generalizability to unseen segmentation tasks.
• The approach is effective across two publicly available skin lesion datasets.
• The framework is model-agnostic and can be combined with different architectures.