Segmentation (src/segment.py)

Handles extracting individual cell crops from raw microscope images.

Architecture

We rely on Cellpose, an advanced instance segmentation framework that generalizes well to new image modalities without extensive fine-tuning.

Primary Functions

• Converts images to grayscale/normalization if required.
• Invokes cyto2 model to handle overlapping structures and cytoplasmic artifacts common in Papanicolaou-stained smears.
• Iterates over generated segmentation masks to slice bounding-box cropped images for classification.

Challenges Solved

Cell clumps present a significant hurdle in cytology. Cellpose's flow-prediction algorithm excels at distinguishing closely packed or overlapping boundaries where traditional watershed or thresholding algorithms fail.