Deploying YOLO for whole-slide mitosis detection: A transparent, auditable, and reproducible pipeline for pathology AI-assisted diagnosis and decision-support tool

Reliable mitotic identification in whole-slide images (WSIs) is essential for prognostic evaluation and tumor grading; yet, many commercial artificial intelligence (AI) pathology platforms function as opaque black boxes with little auditability and, frequently, little expert supervision throughout development.

We developed a transparent and repeatable process for mitosis identification using YOLOv8, designed to scale from full slide inference to patch-level benchmarking. Global non-maximum suppression (NMS), GPU accelerated inference, region of interest-driven and adaptive tiling, and tissue gating are all integrated into the pipeline. In order to facilitate expert validation and traceability, audit-ready outputs include detection CSVs, tile manifests, positive and negative overlays, and heatmaps.

From small clinical slides (26 tiles, 4 detections) to huge benchmarks (TUPAC SP2 with 14,270 tiles, 1754 detections), six typical WSIs were processed. Whereas throughput averaged 78 tiles/s depending on the slide size and tissue dispersion, average confidence scores remained constant between 0.69 and 0.80. The successful generation of all audit outputs attested to its scalability and reproducibility.

Transparency and expert monitoring are given priority in this pipeline, in contrast to many industry systems that place an emphasis on end-to-end automation. Whereas audit outputs enable reviewers to track each detection, explicit recording of positive detection tiles avoid quiet false positives. Evaluation on a small number of WSIs and the lack of benchmarking against annotated entire slide datasets are among the limitations, highlighting the necessity for more extensive and thorough cohort validation. Another significant drawback of this work was the use of a single GPU for inferencing.

This study develops a therapeutically focused, audit-ready process for mitotic figure detection, hotspot identification, and decision-support tool by fusing technical rigor, vetted by experts throughout the process. By demonstrating scalability, reproducibility, and transparency, this work is positioned as a proof-of concept for clinical mitosis detection, multi-GPU/cloud deployment, lab information systems/LIMS connectivity, and eventual use in digital pathology work processes. It also offers a trustworthy and marketable illustration of AI's potential in pathology.