Instance segmentation of fluorescence microscopy cell nuclei using the pre-trained Cellpose model.
This project evaluates segmentation accuracy on 50 paired images and ground-truth masks, reporting metrics such as Precision, Recall, F1, and Dice.
It was completed as part of a technical task under Dr. Madeleine Torcasso at the University of Chicago.

• Goal: Automate segmentation of cell nuclei and evaluate how well a pre-trained model generalizes.
• Dataset: 50 TIFF microscopy images of nuclei, each with paired ground-truth masks.
• Approach:
  1. Normalize raw images to [0,1].
  2. Segment nuclei using the default/general pre-trained Cellpose model.
  3. Match predicted masks to ground-truth masks using IoU ≥ 0.5.
  4. Compute instance segmentation metrics: Precision, Recall, F1, Dice.
  5. Export per-image metrics to CSV and dataset-level averages.

1. Preprocessing: Normalize raw images (0–1 scale).
2. Segmentation: Run Cellpose model → label masks with unique IDs per nucleus.
3. Evaluation:
   • Build an IoU matrix between predicted and ground-truth instances.
   • Greedy matching at IoU ≥ 0.5 → TP, FP, FN.
   • Compute Precision, Recall, F1, Dice.
4. Outputs:
   • CSV with per-image metrics.
   • Dataset averages.
   • Example overlays (for visualization).

Dataset-level averages (IoU ≥ 0.5):

• Precision: 0.824
• Recall: 0.886
• F1 Score: 0.846
• Dice: 0.876

• High recall (0.89): Most true nuclei are detected.
• Slightly lower precision (0.82): Some false positives and merged clumps remain.
• Strong Dice score (0.88): Predicted shapes closely align with ground truth boundaries.

• Model choice: Used the general Cellpose model, not the nuclei-specific weights.
• No fine-tuning: Results could improve if Cellpose were retrained on this dataset.
• Small dataset: Only 50 images; limits generalization conclusions.