Due to the scarcity of annotated data in the medical domain, few-shot learning may be useful for medical image analysis tasks. We design a few-shot learning method using an ensemble of random subspaces for the diagnosis of chest x-rays (CXRs). Our design is computationally efficient and almost 1.8 times faster than method that uses the popular truncated singular value decomposition (t-SVD) for subspace decomposition. The proposed method is trained by minimizing a novel loss function that helps create well-separated clusters of training data in discriminative subspaces. As a result, minimizing the loss maximizes the distance between the subspaces, making them discriminative and assisting in better classification. Experiments on large-scale publicly available CXR datasets yield promising results.

The proposed method is designed for few-shot diagnosis of chest x-ray (CXR) images using an ensemble of random subspaces. It consists of three stages:

1. Feature extraction module (FEM) that extracts visual feature vectors from x-ray images.
2. Subspace embedding module (SEM) that projects the feature vectors into multiple random discriminative subspaces.
3. Final decision module (FDM) that assigns a final class label to an input x-ray image based on the projections.

The SEM creates an ensemble of discriminative subspaces to explore different combinations of visual features obtained from small data samples, resulting in improved classification accuracy. In addition, the proposed method includes a novel loss component that makes these subspaces class discriminative, providing a faster alternative to existing computationally intensive subspace decomposition techniques such as truncated singular value decomposition.

The major contributions of this paper are:

• Proposal of a method for few-shot chest x-ray diagnosis using an ensemble of random subspaces.
• Incorporation of multiple modules along with a novel loss component that improves classification accuracy.
• Providing a faster alternative to existing computationally intensive subspace decomposition techniques.

The proposed method has three modules: feature extraction, subspace embedding, and final decision. Feature extraction extracts visual features, subspace embedding projects feature vectors into multiple subspaces for class assignment, and final decision produces the output label based on these assignments.