TL;DR
This paper introduces PLAB-GAN, a novel semi-supervised deep learning approach for medical image classification that effectively leverages unlabelled data. By combining generative adversarial networks (GANs) with pseudo-labelling and K-means clustering, PLAB-GAN achieves state-of-the-art accuracy on chest X-ray and breast cancer datasets, even with a significantly reduced number of labelled samples.
What Problem Does This Paper Address?
Deep learning (DL) has revolutionized image classification, but its application in medical imaging is hampered by the scarcity of large, labelled datasets. Acquiring and annotating medical images is expensive, time-consuming, and often requires specialized expertise, creating a bottleneck for training robust DL models. This paper addresses this challenge by proposing a method that can learn effectively from a limited amount of labelled data by intelligently incorporating a vast pool of unlabelled medical images.
Methods
The core of the proposed method is the PLAB-GAN (Pseudo-Labelling-based Semi-Supervised Generative Adversarial Network). Here's a breakdown of its key components:
- Semi-Supervised Learning with GANs: The approach utilizes a Generative Adversarial Network (GAN) framework. GANs consist of a generator (which creates synthetic images) and a discriminator (which tries to distinguish real from fake images). In this context, the discriminator is adapted to also act as a classifier for medical images.
- Pseudo-Labelling: This technique assigns "pseudo labels" to unlabelled data. The paper uses K-means clustering on features extracted by a pre-trained ResNet-20 network to group unlabelled images. These clusters, aligned with the labelled data's class centers, are used to infer labels for the unlabelled samples.
- Enhanced Discriminator: The discriminator is modified to classify not only real vs. fake images but also to predict the specific class of a real medical image. This multi-task capability strengthens its learning process.
- Iterative Training: The generator and discriminator are trained iteratively. The generator learns to produce realistic synthetic images that can fool the discriminator, while the discriminator learns to classify both real and synthetic images accurately. This process is refined by using the pseudo-labelled data.
Datasets
The proposed method was evaluated on two publicly available medical imaging datasets:
- ChestX-ray14: A large dataset containing 112,120 chest X-ray images with labels for 14 types of lung diseases. For this study, seven common diseases and a normal class were selected.
- BreakHis: A dataset comprising 7909 breast tissue microscopic images, categorized into benign and malignant subtypes. The images are available at various magnifications.
The paper does not introduce new benchmarks but rather uses these existing, well-established datasets to validate its approach.
Key Insights and Ideas
- Leveraging Unlabelled Data is Crucial: The paper strongly emphasizes that effectively utilizing unlabelled data is key to overcoming the limitations of scarce labelled medical images.
- K-means Clustering for Pseudo-Labels: Using K-means clustering on extracted features provides a robust and simple way to assign pseudo labels, which are then used to guide the GAN training.
- GANs as Classifiers: Adapting the GAN discriminator to perform multi-class classification alongside real/fake discrimination is a powerful way to enhance its learning capacity for medical images.
- Reduced Label Dependency: The method demonstrates that high accuracy can be achieved with significantly fewer labelled samples than traditional supervised methods, making DL more accessible in resource-constrained medical scenarios.
Key Results
The PLAB-GAN method achieved impressive results on both datasets:
- Chest X-ray Classification: Achieved an accuracy of 93.15% using only 400 labelled samples per class, outperforming CNN, PCA+SVM, and SSGAN by a significant margin (18%, 20%, and 16% respectively).
- Breast Cancer Histopathology Classification: Reached an accuracy of 96.87% with as few as 30 labelled samples per class, outperforming existing state-of-the-art methods.
Limitations and Future Work
- Ground Truth Quality: The authors acknowledge that the quality of the initial ground truth labels significantly impacts the pseudo-labelling and subsequent classification performance. Errors in ground truth can propagate and lead to misclassification.
- Future Work: Potential solutions to the ground truth limitation include developing mechanisms to detect out-of-distribution or adversarial samples and exploring novel ambiguity quantification functions to down-weight unreliable training samples. The authors also plan to further investigate strategies to enhance the model's overall performance and robustness.
Summary updated