Non-negative Subspace Feature Representation for Few-shot Learning in Medical Imaging

TL;DR

The paper addresses the challenge of data scarcity in medical imaging by proposing data-based few-shot learning using non-negative subspace representations derived from a fixed pre-trained backbone. It systematically studies NMF and its supervised variants (DNMF, SCNMFS) as dimensionality-reduction tools and compares them to SVD/PCA across 14 medical datasets spanning 11 diseases. The results show that NMF-based subspaces, particularly supervised variants, often outperform SVD, with SCNMFS delivering robust discrimination and improved lesion localization via CAM under few-shot conditions. The work suggests that non-negativity and part-based representations provide practical advantages for medical-imaging inference in low-data regimes and offer a viable alternative to traditional PCA/SVD for subspace-based few-shot learning.

Abstract

Unlike typical visual scene recognition domains, in which massive datasets are accessible to deep neural networks, medical image interpretations are often obstructed by the paucity of data. In this paper, we investigate the effectiveness of data-based few-shot learning in medical imaging by exploring different data attribute representations in a low-dimensional space. We introduce different types of non-negative matrix factorization (NMF) in few-shot learning, addressing the data scarcity issue in medical image classification. Extensive empirical studies are conducted in terms of validating the effectiveness of NMF, especially its supervised variants (e.g., discriminative NMF, and supervised and constrained NMF with sparseness), and the comparison with principal component analysis (PCA), i.e., the collaborative representation-based dimensionality reduction technique derived from eigenvectors. With 14 different datasets covering 11 distinct illness categories, thorough experimental results and comparison with related techniques demonstrate that NMF is a competitive alternative to PCA for few-shot learning in medical imaging, and the supervised NMF algorithms are more discriminative in the subspace with greater effectiveness. Furthermore, we show that the part-based representation of NMF, especially its supervised variants, is dramatically impactful in detecting lesion areas in medical imaging with limited samples.

Figures (6)

• Figure 1: Few-shot learning framework on medical imaging; see the main text for the detailed description. Note that there is no fine-tuning or training process for feature extraction (i.e., the purple block).
• Figure 2: Classification accuracy of the NMF, DNMF, SCNMFS and SVD subspaces on 14 medical datasets with subspace dimensions ranging from 2 to 70. The dataset size is chosen as 600. In the plots, different colours correspond to different methods. Ten random partitions of the training-test set on each of the 14 datasets are conducted. It shows that the supervised NMF, especially SCNMFS, achieves significant improvements over the SVD-based subspace.
• Figure 3: ROC curve of different subspaces (i.e., SVD, NMF, DNMF and SCNMFS) in 5D and 15D subspace representations on the PneumoniaMNIST dataset with the size of 300. The blue, green, yellow and red lines represent the KNN results on SVD, NMF, DNMF and SCNMFS subspaces, respectively. It shows that the performance of SCNMFS is much more stable than others including SVD in both low and high dimensions (i.e., 5D and 15D).
• Figure 4: Sparsity analysis of NMF, DNMF and SCNMFS by investigating the projection matrices $\boldsymbol{U}_{\rm train}^{\Delta}$ and the subspace representations $\boldsymbol{V}_{\rm train}^{\Delta}$ with different subspace dimensions. Dataset PneumoniaMNIST of size 300 is used in this experiment.
• Figure 5: Correlation analysis regarding projection matrices $\boldsymbol{U}_{\rm train}^{\Delta}$ and subspace representations $\boldsymbol{V}_{\rm train}^{\Delta}$ generated by DNMF and SCNMFS with different subspace dimensions. Dataset PneumoniaMNIST is used.