The research team set out to answer a deceptively simple question: which of the newest YOLO-family detectors, and which data balancing strategy, best serves the peculiar demands of mitosis detection? YOLO, short for You Only Look Once, was introduced in 2016 as a unified, real-time approach to object detection. Unlike two-stage detectors such as Faster R-CNN, which first propose candidate regions and then classify them, YOLO models process the entire image in a single forward pass, predicting bounding boxes and class probabilities simultaneously. That speed makes them attractive for digital pathology, where a single whole-slide image can exceed 100,000 by 100,000 pixels and must be tiled into thousands of smaller patches before analysis. The authors evaluated two recent variants: YOLO11n, the nano-sized member of the YOLO11 family, and YOLO12s, a small configuration of the newer YOLO12 architecture, which Ultralytics describes as attention-centric in its design.

The technical distinction between the two models matters. YOLO11n is deliberately lightweight, trading representational capacity for inference speed and a small memory footprint. YOLO12s, while still compact compared with large detection backbones, incorporates attention mechanisms that allow the network to weigh the relative importance of different spatial locations and feature channels when deciding whether a candidate object is a mitotic figure. In histopathology, where mitotic cells are distinguished from look-alikes by subtle cues such as condensed chromatin, absence of a nuclear membrane and the presence of a mitotic spindle, that extra capacity for focusing on fine-grained local detail appears to pay off. The study’s results show YOLO12s consistently outperforming YOLO11n across precision, recall and overall detection accuracy, confirming that the architectural upgrade translates into measurably better detection of these elusive cellular events.

Equally important to the study’s design was the problem of class imbalance. In a typical training set of histopathology patches, non-mitotic cells vastly outnumber mitotic ones, often by orders of magnitude. A naive detector trained on such data can achieve deceptively low error rates simply by never predicting mitosis at all. The authors confronted this challenge by testing two contrasting data balancing strategies: undersampling, in which the abundance of negative examples is reduced to bring the classes closer to parity, and oversampling, in which the rare mitotic examples are repeated or augmented until they are better represented. The choice of strategy proved consequential, interacting with model capacity in ways that shaped the final performance of each detector.

The headline numbers tell the story clearly. YOLO12s achieved an F1-score of 0.9700 when trained with undersampling and 0.9479 with oversampling, while YOLO11n reached 0.9188 and 0.7994 under the same respective conditions. The F1-score, the harmonic mean of precision and recall, is a demanding metric for this task because it penalizes both false positives, in which a dark, hyperchromatic lymphocyte or an apoptotic body is mistaken for a dividing cell, and false negatives, in which a genuine mitotic figure slips through unnoticed. A score of 0.97 means the model’s detections were correct and nearly complete at the same time, a combination that has historically been difficult to achieve in mitosis detection, where early deep learning systems often excelled at one at the expense of the other.

The performance gap between the two balancing strategies is itself instructive. For YOLO12s, undersampling yielded better results than oversampling, suggesting that the model benefits from a cleaner, less redundant training distribution rather than from repeated exposure to duplicated mitotic examples. For the smaller YOLO11n, the gap was far wider, with oversampling dragging the F1-score down to 0.7994. This pattern implies that when a network has limited capacity, flooding it with repetitive minority-class samples can encourage overfitting to the specific appearance of the training mitoses, degrading generalization. The larger effective capacity and attention-based feature selection of YOLO12s appear to buffer the model against this failure mode, allowing it to remain robust under both regimes even though undersampling still gave it the edge.

The study was built on the MIDOG 2021 dataset, a publicly available benchmark created for the Mitosis Domain Generalization Challenge, which the authors note is openly accessible through the challenge website. MIDOG and its successor MIDOG++ were developed specifically to address one of the thorniest problems in computational pathology: domain shift. Whole-slide images are acquired by different scanners, stained in different laboratories and prepared with different protocols, and these variations alter the color, texture and contrast of the tissue in ways that can silently degrade a model trained on data from a single source. By evaluating on a multi-domain dataset, the authors positioned their comparison within a research community effort, documented in Medical Image Analysis, that explicitly tests whether mitosis detectors can generalize beyond the conditions of their training data.

The new work sits within a decade-long lineage of deep learning approaches to mitosis detection. An early landmark came in 2013, when Ciresan and colleagues demonstrated that deep convolutional neural networks could detect mitoses in breast histology images with human-competitive accuracy, winning the ICPR 2012 contest on the task. Subsequent efforts explored a variety of architectures: DeepMitosis combined deep detection, verification and segmentation networks; cascade convolutional models paired with U-Net segmentation were applied to breast cancer slides; and frameworks such as Faster R-CNN, region-based convolutional networks and partially supervised learning were all brought to bear. More recently, MitNet introduced a two-stage deep learning approach on a dedicated whole-slide dataset, and work published in 2024 emphasized robust and efficient detection of mitotic figures across domains. The present study extends this tradition by benchmarking the very latest YOLO generations, which had previously been validated mainly in domains such as agriculture and general object detection rather than in clinical histopathology.

Why does automated mitosis counting matter so much clinically? Breast cancer grading, under widely used schemes, assigns scores based on tubule formation, nuclear atypia and mitotic count, and the mitotic score is often the strongest single predictor of proliferation. Manual counting, however, is laborious and subject to inter-observer variability: pathologists must select a hot spot of high mitotic activity, count dividing cells within a defined field area, and repeat the process across multiple fields, all while fatigued and working against time constraints. Studies dating back to the AMIDA-13 and TUPAC16 grand challenges have shown that algorithmic proliferation estimates from whole-slide images can correlate with tumor grade and patient outcome. A fast, accurate detector such as the one described here could therefore serve as a second reader, flagging hot spots for human review or producing standardized, reproducible mitotic counts that reduce diagnostic variability between institutions.

The authors caution, in effect, that their findings represent a step toward more reliable digital pathology workflows rather than a finished clinical product. Prospective validation on independent cohorts, integration with slide-scanning pipelines and regulatory review would all be required before such a system could support real diagnostic decisions. Nevertheless, the study demonstrates that attention-centric object detection architectures, paired with careful handling of class imbalance, can push mitosis detection performance to a level where the technology becomes genuinely practical. As whole-slide imaging becomes standard in pathology laboratories worldwide, the bottleneck is shifting from image acquisition to image interpretation, and results like these suggest that the newest generation of YOLO models is ready to help clear it.

Subject of Research: Deep learning-based detection of mitotic figures in whole-slide images of breast cancer

Article Title: Detecting mitosis by analyzing whole-slide images of breast cancer using deep learning models

Article References: Iqtaish, N. M., Ali, M. Z., Alorjani, M. S., & Alwadi, M. G. (2026). Detecting mitosis by analyzing whole-slide images of breast cancer using deep learning models. Neural Computing and Applications, 38(19), Article 788. https://doi.org/10.1007/s00521-026-12502-4

Image Credits: AI Generated

DOI: 10.1007/s00521-026-12502-4

Keywords: mitosis detection, breast cancer, whole-slide imaging, deep learning, YOLO12s, YOLO11n, digital pathology, object detection, MIDOG 2021, histopathology, tumor grading, class imbalance