MitoDetect++ is an integrated deep learning pipeline for mitotic phase detection and atypical mitotic classification. Detection (Track 1) utilizes a U-Net-based encoder-decoder architecture with EfficientNetV2-L as a backbone and an attention module, trained using a joint segmentation loss. Classification (Track 2) utilizes a Virchow2 vision transformer, fine-tuned efficiently using Low-Rank Adaptation (LoRA), to minimize resource consumption. It integrates powerful augmentation, focal loss, and group-aware hierarchical 5-fold cross-validation to improve generalization performance and mitigate domain shift. Test-Time Augmentation (TTA) is deployed at inference time to enhance robustness. It achieves a balanced accuracy of 0.892 on the validation domain, highlighting its clinical applicability and cross-task scalability.
