Abstract
In this research, we fine-tune the current stateof-the-art YOLO26x model for automated cochlear hair cell detection. YOLO26, the latest generation of the YOLO family, introduces end-to-end NMS-free inference, Small-Target-Aware Label Assignment (STAL), Progressive Loss Balancing (ProgLoss), and the MuSGD optimizer. For comparative analysis, we also fine-tune three prior YOLO generations (YOLOv10x, YOLO11x, and YOLOv12x) on the same manually annotated cochlear hair cell dataset under identical training conditions (2,000 epochs, 𝟏 𝟎 𝟐 𝟒 × 𝟏 𝟎 𝟐 𝟒 resolution, Ultralytics default hyperparameters). Our YOLO26x model achieves the best overall precision (0.984), recall (0.891), and mAP50 (0.942), with mAP50-95 (0.808) comparable to YOLOv10x (0.810). Notably, YOLO26x also achieves the fastest total inference time (7.1 ms) due to its NMS-free design, compared to 71.1 ms for YOLOv12x. YOLOv12x underperforms all other models despite having the most parameters. To the best of our knowledge, this is the first time YOLO26 has been fine-tuned for cochlear hair cell detection. The code and dataset can be found here: https://github.com/ai-research-2025/DPH₂026