Preprocessing

The preprocessing step prepares a raw SpatialData image for segmentation by performing the following operations:

1. Load morphology image from a SpatialData Zarr file.

   zarr_path = os.path.join(self.zarr_dir, self.zarr_name)
   self.sdata = sd.read_zarr(zarr_path)
   
   img_xr = self.sdata.images[f"{self.fov}_image"]
   img = img_xr.transpose("y", "x", "c").data.compute()
   

2. Extract channels: Membrane (Y), DNA (U).

   channel_indices = [channel_names.index(ch) for ch in channels_to_use]
   img = img[..., channel_indices]
   

3. Normalize contrast (2–98% percentile stretch).

   p2, p98 = np.percentile(img, (2, 98))
   img_stretched = np.clip((img - p2) / (p98 - p2), 0, 1).astype(np.float32)
   

4. Convert to 8-bit RGB (required for Cellpose).

   img_8bit = (img_stretched * 255).astype(np.uint8)
   
   if img_8bit.ndim == 3 and img_8bit.shape[-1] == 2:
       zero_channel = np.zeros_like(img_8bit[..., :1])
       img_8bit = np.concatenate([img_8bit, zero_channel], axis=-1)
   elif img_8bit.ndim == 2:
       img_8bit = np.stack([img_8bit] * 3, axis=-1)
   

5. Downsize to 768×768 (ideal for Cellpose ViT cyto3 on low RAM).

   seg_pil = Image.fromarray(img_8bit, mode="RGB")
   seg_pil.thumbnail(thumbnail_size=(768, 768), Image.Resampling.LANCZOS)
   

6. Save a segmentation-ready .png file.

   if output_path is not None:
       seg_pil.save(output_path)
   
   self.image = seg_pil