TLDR: A new semi-supervised learning framework called SPARSE uses a GAN-based approach with class-conditioned image translation and an ensemble pseudo-labeling technique to achieve high accuracy in medical image classification, even with very few labeled samples (as low as 5 per class). It consistently outperforms existing methods by leveraging real unlabeled images for feature enrichment and combining predictions from multiple models.
Deep learning has brought about significant advancements in medical imaging, offering powerful tools for diagnosis and analysis. However, a major hurdle remains: the scarcity of labeled training data. Obtaining high-quality labeled medical images is challenging due to strict privacy regulations, ethical guidelines, and the need for specialized medical expertise for accurate annotation. This often leads to what researchers call a ‘low-data regime’, where the available labeled images are insufficient for reliably training complex deep neural networks.
Traditional approaches to address this data scarcity, such as unsupervised learning, supervised learning with techniques like transfer learning or data augmentation, and even earlier semi-supervised methods, have their limitations. Unsupervised methods, while utilizing abundant unlabeled data, struggle to ensure that the learned features are clinically relevant. Supervised methods often find their effectiveness diminished when the target medical domain differs significantly from the source data, or they cannot introduce truly new information through augmentation. Recent advancements in Generative Adversarial Networks (GANs) have improved semi-supervised learning, but many existing GAN-based methods still falter when faced with extremely limited labeled data, especially in medical imaging.
Addressing these critical challenges, a new framework named SPARSE (Semi-supervised Pseudo-labeling via Adversarial Representation tranSlation Enhancement) has been introduced. This novel GAN-based semi-supervised learning method is specifically designed to achieve robust classification performance even in scenarios with extremely few labeled samples per class, such as 5 to 50 labeled images.
How SPARSE Works
SPARSE integrates three specialized neural networks: a generator (G), a discriminator (D), and a dedicated classifier (C). Unlike many existing GAN-based methods that generate images from random noise, SPARSE employs a unique image-to-image translation mechanism. This means it modifies existing real, unlabeled images to create new training samples, which helps preserve authentic anatomical features and enriches feature representations more effectively than generating images from scratch.
The framework operates through a dynamic, three-phase training schedule:
- Supervised Training Phase: In this initial phase, all three networks (G, D, and C) are jointly trained using the limited available labeled data. This step is crucial for establishing a foundational understanding and preventing the models from drifting during unsupervised learning.
- Self-Supervised Pre-training Phase: This phase leverages the abundant unlabeled images. It involves an ensemble-based pseudo-labeling block, where the discriminator and classifier combine their confidence-weighted predictions, along with temporal consistency (using an exponential moving average of past predictions), to assign reliable ‘pseudo-labels’ to unlabeled data. Subsequently, the generator uses these pseudo-labels to perform class-conditioned image translation, learning to transform images based on target classes.
- Synthetic Data Enhancement Phase: The translated images, now effectively ‘labeled’ with their target class conditions, are used to further train the classifier. This expands the effective training set, strengthening the classifier’s ability to distinguish between classes.
The training schedule cleverly alternates between the supervised phase (at every epoch) and the combined self-supervised and synthetic data enhancement phase (every few epochs), ensuring stable and efficient use of both labeled and unlabeled data.
Also Read:
- Orchestrating Open-Source AI for Enhanced Medical Diagnosis
- Text-Guided AI Improves Lesion Detection in CT Scans
Impressive Results in Medical Imaging
The effectiveness of SPARSE was rigorously evaluated across eleven diverse MedMNIST datasets, which cover various medical imaging modalities and classification tasks. The method was tested in challenging few-shot settings, ranging from 5-shot (the most difficult scenario) to 50-shot per class.
The results demonstrate that SPARSE, particularly its ensemble configuration (SPARSEens, which combines the discriminator and classifier at inference), consistently outperforms six state-of-the-art GAN-based semi-supervised methods. In the extreme 5-shot setting, SPARSEens achieved statistically significant improvements, showcasing its robust performance even with minimal labeled data. This superiority is largely attributed to its image-to-image translation approach, which maintains the authenticity of medical image features, and its confidence-weighted temporal ensemble technique for generating highly reliable pseudo-labels.
The research also explored the impact of unsupervised training frequency, finding that a moderate alternation (running the unsupervised phase every 10 epochs) consistently yielded optimal performance across all data availability scenarios. This finding provides practical guidance for deploying the method in real-world clinical applications, reducing the need for extensive hyperparameter tuning.
In conclusion, the SPARSE framework offers a practical and powerful solution for medical imaging applications where data annotation costs are prohibitive. By effectively leveraging both labeled and unlabeled data through innovative image translation and ensemble techniques, it enables robust classification performance with as few as five labeled samples per class. For more technical details, you can refer to the full research paper here.
