README.md

metadata

title: Hemocytometer Ground Truth
emoji: 🎯
colorFrom: blue
colorTo: yellow
sdk: gradio
sdk_version: 5.35.0
app_file: app.py
pinned: false

Hemocytometer ground truth

Annotation tool for building ground-truth counts of the 4×4 hemocytometer block. Two tabs: Single image and 4 squares.

Single image

1. Upload one photo.
2. Click the 4 corners of the 4×4 block (any order). Drag a handle to adjust it.
3. Segment. The block is warped to an upright square (≤1024 px) and segmented with the fine-tuned hemocytometer Cellpose-SAM model.
4. Correct the count (see below).

4 squares

1. Upload the four square images at once. They are sorted by filename into Square 1–4.
2. Set each block: click its 4 corners, or press ⌖ Auto and click the block centre to auto-detect the grid. Drag any handle to adjust.
3. Segment all 4. One GPU call per square, as in the main counter. The time from the click until all four squares are ready to annotate is recorded (end-to-end), along with the server's own time.
4. Correct each square on the large canvas below. Switch squares with the buttons or keys 1–4; edits on each square are kept when you switch.
5. Copy the results row (or Copy with header the first time) and paste it into a spreadsheet. Columns: timestamp, files, model_sq1, corrected_sq1, … model_sq4, corrected_sq4, mean_corrected, seg_time_s, server_time_s, min_size_px. Model is what Cellpose counted (after min size and exclusion); corrected is your ground truth. The table also shows the count error, (model − corrected) / corrected.

Correcting the count

• click a cell → remove it (false positive); click again → restore
• click empty space → add a dot (false negative)
• click a dot → delete it
• Shift+click → always add a dot, even on top of a cell (e.g. two cells merged into one mask)
• Ctrl+Z / Ctrl+Shift+Z undo / redo · Space (hold) hide outlines · F fit
• mouse wheel zooms, drag pans

The total is the count of detected objects, minus objects below the min size, minus objects touching the left or bottom 1% exclusion zones, minus removed, plus added dots.

Everything after segmentation runs in the browser, so clicks are instant. The page keeps nothing: copy the results row (or write the count down) before uploading the next image.

Files

app.py (server), gt_canvas.js (annotation canvas), corner_canvas.js (4-corner canvases in the 4-squares tab), grid_seeded.py (grid auto-detection, a copy of the main app's), requirements.txt.

Set hardware to ZeroGPU in the Space settings. To use local weights instead of the Hub file, set CPCC_LOCAL_MODEL=/path/to/weights.npy.