Detecting breast microcalcifications with YOLO — small object detection challenges

1

I’m working on a project to detect breast microcalcifications in mammograms using YOLO (Ultralytics). My dataset consists of high-resolution images (2048×2048) with precise YOLO-format annotations.

Main challenge:

  • Microcalcifications are extremely small (a few pixels) and tend to disappear after resizing to 640×640.

  • I’ve tried CLAHE preprocessing and some augmentations, but recall is still low.

  • Training on full images seems less effective than tiling, but I’m not sure about the best tiling strategy and how to properly recalculate YOLO labels.

Questions to the community:

  1. What image sizes or tiling strategies would you recommend for very small objects in medical imaging?

  2. Have you had success enabling the P2 head (stride=4) in YOLO to improve small object detection?

  3. Which preprocessing steps or hyperparameter settings have worked well in similar (medical or non-medical) projects?

Technical details:

  • Model: YOLOv8n / YOLOv8s (currently testing)

  • Parameters: imgsz=640, mosaic=0.2, conf=0.05, iou=0.4

  • Dataset: mammograms in DICOM converted to PNG, YOLO annotations

  • Environment: Google Colab, T4 GPU

Thanks a lot for your suggestions and shared experiences.