Optimizing Whole-Slide Image Indexing Through Redundancy Reduction

The surge in digital pathology has led to a pressing requirement for effective indexing and retrieval of whole-slide images (WSIs). This demand is further amplified by the rise of generative AI techniques, especially retrieval-augmented generation (RAG), which depend on reliable similarity searches for critical clinical decisions. However, the high costs associated with advanced storage solutions pose challenges for many healthcare facilities in scaling WSI indexing. Therefore, research focusing on reducing storage needs while maintaining retrieval accuracy is vital. We introduce ARReST (Antithetical Redundancy Reduction Strategy), a strategic framework that utilizes redundancy across different tissue types to significantly lower the number of patches that need indexing from each WSI. Rather than just removing duplicates within the same class, ARReST targets antithetical patches—those that minimally aid in distinguishing between classes—and eliminates them from the index. This focused approach effectively compresses the index while preserving both morphological variety and retrieval accuracy. By reducing unnecessary patch representations, ARReST not only minimizes storage requirements but also decreases computational demands and speeds up similarity searches in extensive pathology databases. Tests conducted on the TCGA repository show that ARReST can achieve substantial index compression, with storage savings ranging from 3% to 60%, all while ensuring competitive retrieval performance across various organs. This innovative strategy is poised to facilitate scalable and cost-effective WSI indexing, aligning well with future retrieval-driven clinical AI applications.