A Domain Generalization Algorithm for Mitosis Detection Based on Local Feature Alignment

Mitosis detection is crucial for tumor diagnosis and prediction, yet, facing a domain shift challenge among different pathological images, which results in a poor generalization ability. Here, we introduce a Local Feature Alignment algorithm (LFA) to optimize the model's generalization ability. It maps the local features of the mitotic target to a representation vector space through a linear mapping function. First, we studied the ability of different data augmentation algorithms in improving the model domain generalization performance. Then we constructed YOLO-LFA network using YOLOv5 as a benchmark to achieve the alignment of similar samples from different domains using an improved central loss function. Upon validation with the MIDOG2022 mitosis dataset, the model achieves an average precision of 67.1%, which represents a 0.9% improvement compared to YOLOv5. This underscores the ability of LFA in enhancing the generalization performance of mitosis detection within the domain.