Improved Anomaly Detection in Medical Images via Mean Shift Density Enhancement
Summary: arXiv:2604.19191v1 Announce Type: cross
Abstract
Anomaly detection in medical imaging is essential for identifying rare pathological conditions, particularly when annotated abnormal samples are limited. We propose a hybrid anomaly detection framework that integrates self-supervised representation learning with manifold-based density estimation, a combination that remains largely unexplored in this domain.
Framework Overview
The proposed framework utilizes a two-step process for effective anomaly detection:
- Feature Embedding: Medical images are first embedded into a latent feature space using pretrained, potentially domain-specific, backbones.
- Density Refinement: These representations are then refined via Mean Shift Density Enhancement (MSDE), an iterative manifold-shifting procedure that moves samples toward regions of higher likelihood.
Anomaly Scoring
Anomaly scores are computed using Gaussian density estimation in a PCA-reduced latent space. The Mahalanobis distance is employed to measure deviation from the learned normal distribution. The framework adheres to a one-class learning paradigm, requiring only normal samples for training.
Experimental Validation
Extensive experiments have been conducted on seven medical imaging datasets to validate the efficacy of the proposed framework. The results indicate state-of-the-art performance:
- MSDE achieves the highest Area Under the Curve (AUC) on four datasets.
- It also secures the highest Average Precision (AP) on five datasets, demonstrating its robustness.
- Notably, the framework exhibits near-perfect performance in brain tumor detection, achieving an AUC/AP of 0.981.
Implications for Clinical Practice
These results underscore the potential of the proposed framework as a scalable clinical decision-support tool. Its applications include:
- Early disease detection, facilitating timely intervention.
- Screening in low-label settings where annotated data is scarce.
- Robust deployment across diverse imaging modalities, enhancing clinical workflows.
Conclusion
In summary, the integration of self-supervised representation learning with manifold-based density estimation through Mean Shift Density Enhancement presents a novel approach to anomaly detection in medical imaging. The framework not only achieves remarkable performance metrics but also addresses the challenges posed by limited annotated data, paving the way for improved diagnostic capabilities in clinical settings.