Nuclear Image Analysis: Human vs. AI Benchmark

Welcome to our collaborative research project! This semester, we are benchmarking various image analysis workflows—ranging from manual counting to Deep Learning—to quantify how object overlap (0%–45%) impacts scientific data integrity. You will learn what are the challenges and benefits of each approach.

Your contributions will serve as the foundation for a comparative study on Inter-rater variability and Model Robustness.

---

Project Structure

Each student is assigned a unique ID (student_01 and so on). Your workspace consists of:

• images/: Your assigned 15 TIFF images.
• images/student_XY/masks/: A subfolder for your Phase 4 pixel-wise segmentations.
• student_XY_counts.csv: Your master reporting table. (Copy the template to your folder and replace the XY by your ID)

---

The Reporting Table

Update your .csv file incrementally. Do not change column names or the file structure.

Column	Description
image_id	Filename of the image (e.g., img_01.tif).
manual_count	Points clicked manually in ImageJ/Fiji.
script_own	Results from your custom-tuned Fiji macro.
script_simple	Results from basic global thresholding (Class Workshop).
script_watershed	Results from classic Watershed segmentation (Class Workshop).
script_StarDist	Results using the Pre-trained StarDist model.
YOLO_count	Number of bounding boxes detected (Phase 5 - Optional).

---

Semester Phases & Instructions

Phase 1: Manual Baseline

• Task: Perform manual point counting in Fiji/ImageJ.
• Goal: Establish a "Human Gold Standard" and measure inter-rater variability.
• Note: Some images are hidden duplicates to test your individual consistency (Intra-rater test).

Phase 2: Classic Computer Vision

• Task: Apply global thresholding and Watershed. Compare your own script vs script we make together in class.
• Focus: Identify the "breaking point" where classic algorithms fail to separate overlapping nuclei. Test limits of reproducibility.

Phase 3: Pre-trained StarDist

• Task: Run the pre-trained StarDist model (instance segmentation).
• Focus: Observe how Star-convex polygons handle the 30% and 45% overlap cases compared to Watershed.

Phase 4: Pixel-wise Annotation

• Task: Create precise binary masks for your unique image set.
• Output: Save results as .tif files in your /mask subfolder.
• Format: Name them identically to the source image (e.g., img_01.tif source $\rightarrow$ img_01.tif mask). These will be used to evaluate pixel-level accuracy (Jaccard/Dice indices).
• Focus: Pixel-wise comparison, differences in anotation, challenges in annotation and model training.

Phase 5: YOLO Detection (Optional / Bonus)

• Task: Annotate Bounding Boxes and train a YOLO object detection model.
• Focus: Compare detection-only accuracy against full instance segmentation.

---

Technical Guidelines

1. Decimal Separator: Always use a DOT (.) in your CSVs.
2. Data Integrity: Do not "correct" your manual counts after seeing the AI results. We need to document the raw human error.
3. Git Workflow: * git pull before starting. (Or fork.)
   • git commit -m "Update Phase X for student_XY"
   • git push to your specific branch/folder.

---

Evaluation Metrics

We will be using the following metrics for our final analysis:

• MAE (Mean Absolute Error): For counting precision.
• Bland-Altman Plots: To visualize bias in different methods.
• Jaccard Index: To measure the overlap between your Phase 4 masks and the Ground Truth.

---

Licencing

The original image data and ground truth masks are from "image set BBBC004v1 [Ruusuvuori et al., in Proc. of the 16th European Signal Processing Conference (EUSIPCO-2008), 2008] from the Broad Bioimage Benchmark Collection Ljosa et al., Nature Methods, 2012." This data is licenced udner a Creative Commons Attribution-NonCommercial-ShareAlike 3.0 Unported License by Pekka Ruusuvuori.

Rest of the content is under Creative Commons Attribution 4.0 International licence.

---

Contact: If you find corrupted images or have issues with the StarDist plugin, please open an "Issue" in this repository.