Hash Guided Patch Retrieval for Efficient Medical Anomaly Detection

Anomaly detection becomes really important in the case of scarce annotated data, such as in biomedical applications where abnormal findings are often rare and spatially localized. Among feature-based methods, Patchcore has emerged as a promising solution for Anomaly Detection by learning the distribution of normal images at the feature-patch level using nearest-neighbor retrieval framework. However, to reduce memory and processing expenses, PatchCore uses coreset subsampling, a concept that works well in industrial inspection settings but is problematic in medical image analysis, where uncommon patterns of normal anatomy should not be eliminated. To deal with this, we present a deep position-aware hashing-based patchcore that conducts quick Hashing-based shortlisting and accurate feature-space comparison in place of compressing the memory bank via coreset selection. This technique preserves the representational completeness needed for medical diagnostics, at the expense of a bigger memory bank, while preserving the tractability of nearest-neighbor search across that bank as it expands. This method successfully maintains a balance between diagnostic sensitivity, localization, and clinical scalability under practical memory constraints, making PatchCore-style anomaly detection, quite appropriate for medical diagnostics applications. This method is suitable for large-scale deployment due to efficient retrieval. The code is available at https://github.com/Priyam2323/H_Patchcore/tree/main .