Performance analysis of the YOLO series for object detection: Detection of mitosis cells in histopathology images

Since mitosis cells are one of the most important prognostic indicators used to measure the aggressiveness of cancer, regardless of the type of tissue in which the cancer occurs, it is extremely important to correctly identify mitosis cells on pathology slides. The main motivation of this study is to eliminate the problems of mitosis detection in pathology by computerizing the mitosis detection task and to analyze the potential of YOLO architectures, a series of single-stage object detection models, to create real-time clinical applications that reduce the workload of pathologists. In this research, mitosis cell detection task was performed using popular YOLO models (YOLOv3, YOLOv5, YOLOv7 and YOLOv8) on MIDOG 2022 dataset containing images from 5 different cancer types. In the evaluations, it was observed that YOLOv7 and YOLOv8 architectures outperformed the other models in correctly detecting mitosis cells. The most successful model was YOLOv8, which achieved an impressive Recall value of 89.1%. It was observed that YOLOv7 and YOLOv8 architectures trained with cancer types with higher representation and relatively more balanced data yielded equivalent results. On the other hand, cancer types represented in a relatively lower rate in the dataset and with problems in image quality, YOLOv8 architecture produced more robust results for mitosis detection task in histopathological images.