Abstract
-Patch extraction from gigapixel whole-slide images, typically guided by tissue detection methods, forms the backbone of computational pathology workflows, - and remains a major computational bottleneck. Here we present AtlasPatch, an efficient and scalable slide preprocessing framework designed to enable accurate tissue - detection and high-throughput patch extraction with minimal computational overhead. To ensure robust and generalizable slide processing, we curated and - semi-manually annotated a diverse dataset of approximately 35,000 whole-slide image thumbnails spanning multiple cohorts, tissue types, and organ systems. Using this - corpus, we implement selective finetuning of the normalization layers of the Segment-Anything2 model for efficient thumbnail-level segmentation. This strategy ensures - fast processing while achieving segmentation accuracy comparable to—or exceeding—that of existing methods. From the thumbnail masks, tissue coordinates are efficiently - extrapolated to full-resolution slides for precise patch extraction. The entire pipeline is parallelized for both CPU and GPU execution to maximize throughput. We assess - AtlasPatch across segmentation accuracy, computational complexity, and downstream multiple-instance learning performance, showing consistent predictive power with - state-of-the-art methods while operating at a fraction of their computational cost.
- -
+ Abstract
+Whole-slide image (WSI) preprocessing, typically comprising tissue detection followed by patch extraction, is foundational to AI-driven + and image-based computational pathology workflows. This remains a major computational bottleneck as existing tools either rely on inacurate + heuristic thresholding for tissue detection, or adopt AI-based approaches trained on limited-diversity data that operate at the patch level, + incurring substantial computational complexity. We present AtlasPatch, an efficient and scalable slide preprocessing framework for accurate + tissue detection and high-throughput patch extraction with minimal computational overhead. AtlasPatch’s tissue detection module is trained + on a heterogeneous and semi-manually annotated dataset of $\sim$35,000 WSI thumbnails, using efficient fine-tuning of the Segment Anything2 + model. The tool extrapolates tissue masks from thumbnails to full-resolution slides to extract patch coordinates at user-specified magnifications, + with options to stream patches directly into commonly used image encoders for embedding generation or export patch images for storage, all efficiently + parallelized across CPUs and GPUs to maximize throughput. We assess AtlasPatch across segmentation accuracy, computational complexity, and downstream + multiple-instance learning, matching state-of-the-art performance while operating at a fraction of their computational cost.
+ +
+ 
+ 
+ 
+ 
+ 
+ 
+