Hybrid sizing: detect@0.3 + measure@0.5 core — fixes 10% size bias

inference.py

@@ -261,26 +261,41 @@ def analyze_image(
     cell_prob = 1.0 / (1.0 + np.exp(-out[2]))  # sigmoid
     dist_transform = np.maximum(out[3], 0)
 
-    # Instance segmentation
+    # Instance segmentation (detect at prob_threshold for good recall)
     masks = segment_instances(cell_prob, dist_transform, prob_threshold, min_size_px)
 
-    # Measure bubbles from instance masks
+    # Measure bubbles from instance masks.
+    # KEY: size each bubble using only the high-confidence core (>0.5) to
+    # avoid threshold-dependent bloating. Detection at 0.3 finds bubbles;
+    # the 0.5 interior gives accurate area → diameter.
+    SIZE_THRESHOLD = 0.5
     bubbles = []
     num_rejected = 0
 
     for region in regionprops(masks):
-        circ = (4 * np.pi * region.area) / (region.perimeter ** 2 + 1e-8)
-        major = getattr(region, "axis_major_length", 0) or getattr(region, "major_axis_length", 0)
-        minor = getattr(region, "axis_minor_length", 0) or getattr(region, "minor_axis_length", 0)
-        ar = major / (minor + 1e-8)
+        inst_mask = (masks == region.label)
+
+        # Sizing: use the high-confidence interior if available
+        core = inst_mask & (cell_prob > SIZE_THRESHOLD)
+        core_area = int(core.sum())
+        if core_area >= 5:
+            sizing_area = core_area
+        else:
+            sizing_area = region.area  # fallback for faint detections
 
-        diam_px = getattr(region, "equivalent_diameter_area", 0) or getattr(region, "equivalent_diameter", 0)
+        diam_px = np.sqrt(sizing_area / np.pi) * 2
         diam_um = diam_px * scale_um_per_pixel
         radius_um = diam_um / 2
         vol = (4 / 3) * np.pi * radius_um ** 3
 
-        # Distance-transform radius (peak within this instance)
-        inst_mask = (masks == region.label)
+        # Shape QC from the full detection mask (not the core)
+        circ = (4 * np.pi * region.area) / (region.perimeter ** 2 + 1e-8)
+        major = getattr(region, "axis_major_length", 0) or getattr(region, "major_axis_length", 0)
+        minor = getattr(region, "axis_minor_length", 0) or getattr(region, "minor_axis_length", 0)
+        ar = major / (minor + 1e-8)
+
+        # Distance-transform peak (for reference — note: student DT is normalised,
+        # not in absolute pixels, so this is relative only)
         dt_vals = dist_transform[inst_mask]
         radius_dt = float(dt_vals.max()) if len(dt_vals) > 0 else 0.0
 
@@ -296,7 +311,7 @@ def analyze_image(
             bubble_id=region.label,
             diameter_um=diam_um,
             volume_um3=vol,
-            area_px=region.area,
+            area_px=sizing_area,
             centroid_x=region.centroid[1],
             centroid_y=region.centroid[0],
             radius_dt_px=radius_dt,