lavlab-shell 0.3.1__tar.gz → 0.4.0__tar.gz

{lavlab_shell-0.3.1 → lavlab_shell-0.4.0}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: lavlab-shell
-Version: 0.3.1
+Version: 0.4.0
 Summary: SHELL Highlights Epithelium and Lumen Locations — whole-slide H&E segmentation
 Project-URL: Documentation, https://github.com/laviolette-lab/shell#readme
 Project-URL: Issues, https://github.com/laviolette-lab/shell/issues

{lavlab_shell-0.3.1 → lavlab_shell-0.4.0}/src/shell/__about__.py

@@ -3,4 +3,4 @@
 # SPDX-License-Identifier: MIT
 """Version information for shell."""
 
-__version__ = "0.3.1"
+__version__ = "0.4.0"

{lavlab_shell-0.3.1 → lavlab_shell-0.4.0}/src/shell/cli.py

@@ -105,7 +105,7 @@ def build_parser() -> argparse.ArgumentParser:
     infer_p.add_argument(
         "--target-mpp",
         type=float,
-        default=1.0,
+        default=2.0,
         help="Desired output resolution (um/pixel).",
     )
     infer_p.add_argument(
@@ -125,6 +125,51 @@ def build_parser() -> argparse.ArgumentParser:
         default=None,
         help="Optional: also save the intermediate EHO image.",
     )
+    infer_p.add_argument(
+        "--save-raw",
+        type=str,
+        default=None,
+        help=(
+            "Optional: save raw model predictions as a 3-band uint8 image "
+            "(band 0 = inner / lumen, band 1 = outer / epithelium, "
+            "band 2 = background), scaled to 0/255. "
+            "Saved at the original input resolution. "
+            "Useful for debugging or re-running post-processing offline."
+        ),
+    )
+    infer_p.add_argument(
+        "--profile",
+        type=str,
+        default="best_effort",
+        choices=["best_effort", "precise", "sensitive"],
+        help=(
+            "Post-processing filter profile. "
+            "'best_effort' (default) balances sensitivity and precision; "
+            "'precise' is more conservative; 'sensitive' keeps more predictions."
+        ),
+    )
+    infer_p.add_argument(
+        "--mode",
+        type=str,
+        default="wsi",
+        choices=["wsi", "biopsy", "tile"],
+        help=(
+            "Post-processing mode. "
+            "'wsi' (default): full pipeline with tissue restriction and urethra detection. "
+            "'biopsy': tissue restriction but no urethra detection. "
+            "'tile': no tissue mask/urethra; reflect-pads predictions before morphological ops. "
+            "Use 'tile' for individual image tiles that lack surrounding context."
+        ),
+    )
+    infer_p.add_argument(
+        "--tile-pad",
+        type=int,
+        default=None,
+        help=(
+            "Reflect-padding in pixels applied on each side in tile mode. "
+            "Defaults to 50%% of the shorter output dimension when not set."
+        ),
+    )
     infer_p.add_argument(
         "--device",
         type=str,
@@ -156,7 +201,7 @@ def build_parser() -> argparse.ArgumentParser:
     omero_p.add_argument(
         "--target-mpp",
         type=float,
-        default=1.0,
+        default=2.0,
         help="Desired output resolution (um/pixel)",
     )
     omero_p.add_argument(
@@ -252,6 +297,17 @@ def build_parser() -> argparse.ArgumentParser:
             "sessions. Default 4."
         ),
     )
+    omero_p.add_argument(
+        "--profile",
+        type=str,
+        default="best_effort",
+        choices=["best_effort", "precise", "sensitive"],
+        help=(
+            "Post-processing filter profile. "
+            "'best_effort' (default) balances sensitivity and precision; "
+            "'precise' is more conservative; 'sensitive' keeps more predictions."
+        ),
+    )
 
     return parser
 
@@ -303,6 +359,10 @@ def main(argv: list[str] | None = None) -> int:
             target_mpp=args.target_mpp,
             mpp=args.mpp,
             save_eho=args.save_eho,
+            save_raw=args.save_raw,
+            profile=args.profile,
+            mode=args.mode,
+            tile_pad=args.tile_pad,
             device=args.device,
         )
 
@@ -351,6 +411,7 @@ def main(argv: list[str] | None = None) -> int:
                 min_tissue_frac=args.min_tissue_frac,
                 prefetch_depth=args.prefetch_depth,
                 num_fetch_workers=args.fetch_workers,
+                profile=args.profile,
             )
             log.debug("DIAG: infer_omero_wsi returned normally")
         except BaseException as e:

{lavlab_shell-0.3.1 → lavlab_shell-0.4.0}/src/shell/infer_omero_wsi.py

@@ -1292,6 +1292,7 @@ def infer_omero_wsi(
     min_tissue_frac: float = 0.01,
     prefetch_depth: int = 8,
     num_fetch_workers: int = 4,
+    profile: str = "best_effort",
 ) -> np.ndarray:
     """Tile-based OMERO inference pipeline.
 
@@ -1633,6 +1634,15 @@ def infer_omero_wsi(
         )
         tiles_tissue = len(tile_schedule)
 
+        # Upscale tissue mask to full resolution for post-processing.
+        th_h, th_w = tissue_mask.shape
+        y_idx = np.clip(
+            (np.arange(out_h) * th_h / out_h).astype(np.int64), 0, th_h - 1,
+        )
+        x_idx = np.clip(
+            (np.arange(out_w) * th_w / out_w).astype(np.int64), 0, th_w - 1,
+        )
+        tissue_mask_full = tissue_mask[y_idx[:, None], x_idx[None, :]]
         del tissue_mask
         gc.collect()
 
@@ -1689,7 +1699,10 @@ def infer_omero_wsi(
             save_eho=save_eho,
             output_path=output_path,
             tile_overlap=tile_overlap,
+            roi_size=inference_tile_size,
             sw_overlap=sw_overlap,
+            tissue_mask_full=tissue_mask_full,
+            profile=profile,
             num_fetch_workers=num_fetch_workers,
         )
 
@@ -1736,7 +1749,10 @@ def _run_pipeline(
     save_eho: str | None,
     output_path: str,
     tile_overlap: int,
+    roi_size: int,
     sw_overlap: float,
+    tissue_mask_full: np.ndarray | None = None,
+    profile: str = "best_effort",
     num_fetch_workers: int = 4,
 ) -> np.ndarray:
     """Producer/consumer pipeline: parallel fetch+EHO → inference in main.
@@ -1818,7 +1834,10 @@ def _run_pipeline(
         except Exception:
             logger.exception("Failed to start OMERO fetch thread; continuing.")
 
-        pred = np.zeros((out_h, out_w), dtype=np.uint8)
+        # Allocate canvases for inner/outer predictions and hematoxylin.
+        inner_pred = np.zeros((out_h, out_w), dtype=bool)
+        outer_pred = np.zeros((out_h, out_w), dtype=bool)
+        hematoxylin_full = np.zeros((out_h, out_w), dtype=np.uint8)
 
         eho_canvas: np.ndarray | None = None
         if save_eho:
@@ -1838,24 +1857,34 @@ def _run_pipeline(
             if eho_canvas is not None:
                 eho_canvas[oy0 : oy0 + oh, ox0 : ox0 + ow] = tile_eho
 
-            tile_pred = run_inference(
+            # Store hematoxylin channel (ch1) for nuclei segmentation.
+            hematoxylin_full[oy0 : oy0 + oh, ox0 : ox0 + ow] = tile_eho[:, :, 1]
+
+            inner_tile, outer_tile = run_inference(
                 tile_eho,
                 model,
                 device,
+                roi_size=(roi_size, roi_size),
                 overlap=sw_overlap,
+                return_raw=True,
             )
             del tile_eho
 
             # Write only the centre-cropped keep region to avoid seams.
-            ph, pw = tile_pred.shape[:2]
+            ph, pw = inner_tile.shape[:2]
             ky1 = min(keep_y0 + keep_h, ph)
             kx1 = min(keep_x0 + keep_w, pw)
-            kept = tile_pred[keep_y0:ky1, keep_x0:kx1]
             out_y0 = oy0 + keep_y0
             out_x0 = ox0 + keep_x0
-            ch, cw = kept.shape[:2]
-            pred[out_y0 : out_y0 + ch, out_x0 : out_x0 + cw] = kept
-            del tile_pred, kept
+            ch = ky1 - keep_y0
+            cw = kx1 - keep_x0
+            inner_pred[out_y0 : out_y0 + ch, out_x0 : out_x0 + cw] = (
+                inner_tile[keep_y0:ky1, keep_x0:kx1]
+            )
+            outer_pred[out_y0 : out_y0 + ch, out_x0 : out_x0 + cw] = (
+                outer_tile[keep_y0:ky1, keep_x0:kx1]
+            )
+            del inner_tile, outer_tile
 
             processed += 1
             if processed % 25 == 0:
@@ -1873,7 +1902,22 @@ def _run_pipeline(
         if torch.cuda.is_available():
             torch.cuda.empty_cache()
 
-        return _save_results(pred, eho_canvas, save_eho, output_path)
+        # ── Post-processing ──────────────────────────────────────────
+        from shell.post_process import post_process
+
+        logger.info("Running post-processing (profile=%s) …", profile)
+        label_map = post_process(
+            inner_pred,
+            outer_pred,
+            tissue_mask_full if tissue_mask_full is not None else np.ones((out_h, out_w), dtype=bool),
+            hematoxylin_full,
+            profile_name=profile,
+            verbose=True,
+        )
+        del inner_pred, outer_pred, hematoxylin_full, tissue_mask_full
+        gc.collect()
+
+        return _save_results(label_map, eho_canvas, save_eho, output_path)
 
     finally:
         fetch_thread.join(timeout=30)