clarkrinker/yolov11n-nft-detector

Model Card for Model ID

A Yolo 11 Object detector for Neurofibrillary tangles.

Model is trained from Dataset provided by Vizcarra et al. Citation:

Vizcarra JC, Pearce TM, Dugger BN, Keiser MJ, Gearing M, Crary JF, Kiely EJ, Morris M, White B, Glass JD, Farrell K, Gutman DA. Toward a generalizable machine learning workflow for neurodegenerative disease staging with focus on neurofibrillary tangles. Acta Neuropathol Commun. 2023 Dec 18;11(1):202.

DOI: 10.1186/s40478-023-01691-x PubMed ID (PMID): 38110981 PubMed Central ID (PMCID): PMC10726581

Model Details

• This Model was trained on immunohistochemically stained slides of brain matter at a resolution of 0.25 microns per pixels.
• Slides are tiled into 1280x1280 segments before being fed to the detector.
• Out of band testing appears to produce good results at .5049 microns per pixel

Example Ouput

Usage

Generating Tiles

This model expects 1280x1280 sized images. If you are working with a whole slide like an .svs file you can generate tiles using openslide:

tile_width=1280
tile_height=1280
svsfile = "myslide.svs"

slide = openslide.OpenSlide(svsfile)
slide_width, slide_height = slide.level_dimensions[level]
for y in range(0, slide_height, tile_height):
  for x in range(0, slide_width, tile_width):
      region = slide.read_region((x, y), level, (tile_width, tile_height)).convert("RGB")
      filename = f"segments/segment_{x}_{y}.png"
      region.save(filename)

With The Yolo Command Line

# pip install ultralytics
yolo predict model=.\yolo11n-nft-detector.pt source=some_tile.png

With The Yolo API

from ultralytics import YOLO

model_path = "yolo11n-nft-detector.pt"
image_path = "some_tile.png"
model = YOLO(model)
res = model(image_path)

Training Results

🎯 Final Model Performance Summary Overall Metrics:

• mAP50: 80.8%

• mAP50-95: 62.3%

• Precision: 72.8% -

• Recall: 74.4% -

• Class Performance:

• Pretangle (Type 0): 71.8% mAP50, 66.3% recall

• Tangle (Type 1): 89.7% mAP50, 82.4% recall

• Speed: 1.2ms inference - 4080 Super