OCSVM-guided method detects subtle brain lesions other AI misses

Unsupervised anomaly detection (UAD) is critical when labeled anomalies are scarce, but existing methods often reconstruct anomalies too well (reconstruction-based) or learn suboptimal feature spaces (decoupled density estimators). To address this, Nicolas Pinon and colleagues from MYRIAD propose a tight coupling of representation learning with a One-Class SVM (OCSVM) through a custom loss that directly optimizes latent features to align with the OCSVM decision boundary. This avoids surrogate objectives and kernel restrictions, allowing the model to learn a more expressive and robust feature space. The model is tested on two challenging tasks: a corrupted MNIST benchmark (MNIST-C) and a brain MRI lesion detection task. Notably, it targets small, non-hyperintense lesions at the voxel level—a more clinically relevant scenario than typical image-level detection of large hyperintense lesions.

The method demonstrates strong robustness to domain shifts, including corruption types in MNIST-C and texture or age variations in MRI data. Compared to state-of-the-art UAD methods, it achieves superior performance on both benchmarks, particularly in detecting subtle anomalies that other models miss. The researchers have released the source code, enabling reproducibility and further experimentation. This work has significant implications for medical imaging, where early detection of small lesions can improve diagnosis and treatment planning. By directly aligning feature learning with the anomaly detection objective, the approach offers a principled route to more reliable unsupervised anomaly detectors in real-world applications.