fix: preserve dark necrotic tissue connected to foot, only remove isolated dark blobs

src/segmentation.py

@@ -261,16 +261,20 @@ def postprocess_segmentation(
     classmap: np.ndarray,
     img_bgr: np.ndarray,
     min_foot_ratio: float = 0.01,
-    dark_l_threshold: float = 20.0,
+    dark_l_threshold: float = 15.0,
 ) -> np.ndarray:
     """Post-process segmentation to remove noise and dark-background misclassifications.
 
     Steps:
         1. Keep only the largest connected component of foreground (foot+peri+ulcer).
            Small disconnected blobs are reclassified as background.
-        2. Exclude very dark pixels (L* < threshold) that are far from the main
-           foot region — these are likely background/shadow, not skin or wound.
+        2. Exclude very dark pixels ONLY if they are NOT part of the main connected
+           foreground component. Dark necrotic tissue connected to the foot is preserved.
         3. Morphological opening to clean noisy edges.
+
+    Key design: necrotic tissue (very dark, low L*) that is connected to the main
+    foot region is NEVER removed — only isolated dark blobs disconnected from the
+    foot are reclassified as background.
     """
     h, w = classmap.shape
     cleaned = classmap.copy()
@@ -279,11 +283,13 @@ def postprocess_segmentation(
     foreground = (cleaned > 0).astype(np.uint8)
     num_labels, labels, stats, _ = cv2.connectedComponentsWithStats(foreground, connectivity=8)
 
-    if num_labels > 2:
-        # Find the largest foreground component (label 0 is background in connectedComponents)
-        # stats[:, cv2.CC_STAT_AREA] gives area of each label
-        areas = stats[1:, cv2.CC_STAT_AREA]  # Skip background (label 0)
-        largest_label = np.argmax(areas) + 1  # +1 because we skipped label 0
+    # Track which pixels belong to the main component
+    main_component_mask = np.zeros((h, w), dtype=bool)
+
+    if num_labels > 1:
+        areas = stats[1:, cv2.CC_STAT_AREA]
+        largest_label = np.argmax(areas) + 1
+        main_component_mask = (labels == largest_label)
         min_area = h * w * min_foot_ratio
 
         for label_id in range(1, num_labels):
@@ -291,26 +297,21 @@ def postprocess_segmentation(
                 continue
             component_area = stats[label_id, cv2.CC_STAT_AREA]
             if component_area < min_area:
-                # Small disconnected blob → background
                 cleaned[labels == label_id] = 0
+    else:
+        main_component_mask = foreground.astype(bool)
 
-    # Step 2: Dark pixel exclusion
-    # Very dark pixels that ended up as foreground are likely background/shadow
+    # Step 2: Dark pixel exclusion — ONLY for pixels NOT in the main component
+    # Necrotic tissue connected to the foot is preserved regardless of darkness
     lab = cv2.cvtColor(img_bgr, cv2.COLOR_BGR2Lab).astype(np.float32)
     l_channel = lab[:, :, 0] * (100.0 / 255.0)
 
     dark_mask = l_channel < dark_l_threshold
     is_foreground = cleaned > 0
 
-    # Only reclassify dark pixels that are NOT adjacent to the main foot core
-    # Strategy: dilate the non-dark foreground region, dark pixels outside it → background
-    core_foreground = (is_foreground & ~dark_mask).astype(np.uint8)
-    if np.any(core_foreground):
-        kernel = cv2.getStructuringElement(cv2.MORPH_ELLIPSE, (31, 31))
-        core_dilated = cv2.dilate(core_foreground, kernel).astype(bool)
-        # Dark foreground pixels NOT touching the core → background
-        dark_isolated = dark_mask & is_foreground & ~core_dilated
-        cleaned[dark_isolated] = 0
+    # Only remove dark pixels that are foreground but NOT in the main component
+    dark_isolated = dark_mask & is_foreground & ~main_component_mask
+    cleaned[dark_isolated] = 0
 
     # Step 3: Morphological opening per foreground class to clean edges
     kernel_small = cv2.getStructuringElement(cv2.MORPH_ELLIPSE, (5, 5))