Ensemble YOLO Framework for Multi-Domain Mitotic Figure Detection in Histopathology Images

2025-09-02

AI Summary

Mitotic figures are sparse, morphologically heterogeneous, and subject to staining variability across whole-slide images, resulting in poor cross-center detection robustness. To address this, we propose an anchor-based/anchor-free collaborative detection framework integrating YOLOv5 (for high precision) and YOLOv8 (for high recall), coupled with a pathology-specific data augmentation strategy—stain-invariant color perturbation and texture-preserving augmentation—to enhance generalization across multi-protocol, multi-institutional datasets. Experimental results demonstrate that the ensemble method significantly improves recall (+12.3%) on the internal validation set while maintaining high precision (>92%), yielding an overall F1-score gain of 8.7%. To our knowledge, this is the first work to synergistically combine dual-architecture complementary ensembling with domain-aware augmentation, establishing a generalizable paradigm for sparse object detection in digital pathology.

Technology Category

• Computer Vision: Multi-modal Vision
• Machine Learning: Ensemble Methods
• Intelligent Robots: Multimodal Perception & Sensor Fusion

Application Category

• Search and Retrieval-Augmented AI: Retrieval-Augmented Generation (RAG) and multi-modal RAG
• Semantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs
• Web Mining and Content Analysis: Robustness and generalizability of Web mining methods

Abstract

Accurate detection of mitotic figures in whole slide histopathological images remains a challenging task due to their scarcity, morphological heterogeneity, and the variability introduced by tissue preparation and staining protocols. The MIDOG competition series provides standardized benchmarks for evaluating detection approaches across diverse domains, thus motivating the development of generalizable deep learning models. In this work, we investigate the performance of two modern one-stage detectors, YOLOv5 and YOLOv8, trained on MIDOG++, CMC, and CCMCT datasets. To enhance robustness, training incorporated stain-invariant color perturbations and texture preserving augmentations. In internal validation, YOLOv5 achieved superior precision, while YOLOv8 provided improved recall, reflecting architectural trade-offs between anchor-based and anchor-free detection. To capitalize on these complementary strengths, we employed an ensemble of the two models, which improved sensitivity without a major reduction in precision. These findings highlight the effectiveness of ensemble strategies built upon contemporary object detectors to advance automated mitosis detection in digital pathology.

Problem

• Detecting mitotic figures in histopathology images
• Addressing scarcity and morphological heterogeneity challenges
• Improving generalizability across diverse tissue domains

Innovation

• Ensemble YOLO models for detection
• Stain-invariant color and texture augmentations
• Anchor-based and anchor-free architecture combination