PathoSCOPE: Few-Shot Pathology Detection via Self-Supervised Contrastive Learning and Pathology-Informed Synthetic Embeddings

TL;DR

PathoSCOPE tackles the problem of pathology detection with minimal non-pathological data by introducing a few-shot unsupervised framework that leverages Global-Local Contrastive Learning and pathology-informed embedding synthesis. The method uses a prototypical anchor bank to anchor normal anatomy and a PiEG module to generate Global Pathological Embeddings through gradient-guided perturbations, enabling robust separation of normal and pathological regions. Evaluated on BraTS2020 and ChestXray8, PathoSCOPE achieves state-of-the-art performance among unsupervised methods while maintaining high throughput and low computational cost. The work demonstrates practical potential for rapid, data-efficient clinical deployment, though it notes limitations in anatomical coherence of synthetic pathologies and points to decoder-based improvements as future work.

Abstract

Unsupervised pathology detection trains models on non-pathological data to flag deviations as pathologies, offering strong generalizability for identifying novel diseases and avoiding costly annotations. However, building reliable normality models requires vast healthy datasets, as hospitals' data is inherently biased toward symptomatic populations, while privacy regulations hinder the assembly of representative healthy cohorts. To address this limitation, we propose PathoSCOPE, a few-shot unsupervised pathology detection framework that requires only a small set of non-pathological samples (minimum 2 shots), significantly improving data efficiency. We introduce Global-Local Contrastive Loss (GLCL), comprised of a Local Contrastive Loss to reduce the variability of non-pathological embeddings and a Global Contrastive Loss to enhance the discrimination of pathological regions. We also propose a Pathology-informed Embedding Generation (PiEG) module that synthesizes pathological embeddings guided by the global loss, better exploiting the limited non-pathological samples. Evaluated on the BraTS2020 and ChestXray8 datasets, PathoSCOPE achieves state-of-the-art performance among unsupervised methods while maintaining computational efficiency (2.48 GFLOPs, 166 FPS).

Figures (5)

• Figure 1: Overall architecture of the PathoSCOPE
• Figure 2: Comparison of t-SNE for synthetic embeddings on ChestXray8 dataset. The pathology-informed embeddings generated by PathoSCOPE more closely resemble real pathological embeddings than GLASS synthetic embeddings.
• Figure 3: Visualization of BraTS2020 Heatmap Prediction
• Figure 4: Comparison of t-SNE for (a) GLASS-Hypersphere, (b) GLASS-Manifold, and (c) PathoSCOPE.
• Figure 5: Evaluation of the perturbation strength ($\eta$) on BraTS2020 and ChestXrat8 datasets.