Lower SAM3 detection thresholds to fix empty masks issue

Co-authored-by: matheus-rech <71096630+matheus-rech@users.noreply.github.com>

app.py

@@ -98,15 +98,15 @@ else:
             processor = None
 
 @spaces.GPU(duration=60)
-def run_sam3_inference(pil_image, prompt_text, threshold=0.5, mask_threshold=0.5):
+def run_sam3_inference(pil_image, prompt_text, threshold=0.1, mask_threshold=0.0):
     """
     Run SAM 3 inference - matching official implementation from akhaliq/sam3.
     
     Args:
         pil_image: PIL Image to segment
         prompt_text: Text prompt for segmentation
-        threshold: Detection threshold (higher = fewer detections)
-        mask_threshold: Mask threshold (higher = sharper masks)
+        threshold: Detection threshold (higher = fewer detections, default 0.1 for medical images)
+        mask_threshold: Mask threshold (higher = sharper masks, default 0.0 for medical images)
     
     Returns:
         results dict with 'masks' and 'scores' as numpy arrays or lists, or None if failed
@@ -444,7 +444,8 @@ def process_medical_image(image_file, prompt_text, modality, window_type, return
             pil_image = Image.fromarray(img_uint8.astype(np.uint8))
 
         # Run SAM 3 Inference - using helper function matching official implementation
-        results = run_sam3_inference(pil_image, prompt_text, threshold=0.5, mask_threshold=0.5)
+        # Lower thresholds for medical images to ensure detections are not filtered out
+        results = run_sam3_inference(pil_image, prompt_text, threshold=0.1, mask_threshold=0.0)
         
         if results is None:
             return None
@@ -639,7 +640,8 @@ def process_medical_image_enhanced(image_file, prompt_text, modality, window_typ
         pil_image = enhancer.enhance(contrast)
 
         # Run SAM 3 Inference - using helper function matching official implementation
-        results = run_sam3_inference(pil_image, prompt_text, threshold=0.5, mask_threshold=0.5)
+        # Lower thresholds for medical images to ensure detections are not filtered out
+        results = run_sam3_inference(pil_image, prompt_text, threshold=0.1, mask_threshold=0.0)
         
         if results is None:
             return None
@@ -966,7 +968,8 @@ def process_with_point_prompt(image_file, point_x, point_y, modality, window_typ
         prompt_text = f"segment region at point"
         
         # Process with SAM 3 - using helper function
-        results = run_sam3_inference(pil_image, prompt_text, threshold=0.5, mask_threshold=0.5)
+        # Lower thresholds for medical images to ensure detections are not filtered out
+        results = run_sam3_inference(pil_image, prompt_text, threshold=0.1, mask_threshold=0.0)
         
         final_mask = None
         if results and 'masks' in results and results['masks'] is not None:
@@ -1062,7 +1065,8 @@ def process_with_box_prompt(image_file, x1, y1, x2, y2, modality, window_type, c
         prompt_text = "segment region in bounding box"
         
         # Process with SAM 3 - using helper function
-        results = run_sam3_inference(pil_image, prompt_text, threshold=0.5, mask_threshold=0.5)
+        # Lower thresholds for medical images to ensure detections are not filtered out
+        results = run_sam3_inference(pil_image, prompt_text, threshold=0.1, mask_threshold=0.0)
         
         final_mask = None
         if results and 'masks' in results and results['masks'] is not None:
@@ -1146,7 +1150,8 @@ def process_multi_mask(image_file, prompt_text, modality, window_type, num_masks
             prompt_text = "brain"
         
         # Process with SAM 3 - using helper function
-        sam_results = run_sam3_inference(pil_image, prompt_text, threshold=0.5, mask_threshold=0.5)
+        # Lower thresholds for medical images to ensure detections are not filtered out
+        sam_results = run_sam3_inference(pil_image, prompt_text, threshold=0.1, mask_threshold=0.0)
         
         results = []
         mask_info = []
@@ -1553,7 +1558,8 @@ def automatic_mask_generator(image_file, modality, window_type,
             
             try:
                 # Process with SAM 3 - using helper function
-                sam_results = run_sam3_inference(pil_image, prompt, threshold=0.5, mask_threshold=0.5)
+                # Lower thresholds for medical images to ensure detections are not filtered out
+                sam_results = run_sam3_inference(pil_image, prompt, threshold=0.1, mask_threshold=0.0)
                 
                 if sam_results and 'masks' in sam_results and sam_results['masks'] is not None:
                     masks = sam_results['masks']  # List of mask tensors
@@ -1732,7 +1738,8 @@ def process_with_advanced_transforms(image_file, prompt_text, modality, window_t
             prompt_text = "brain"
         
         # Process with SAM 3 - using helper function
-        results = run_sam3_inference(pil_image, prompt_text, threshold=0.5, mask_threshold=0.5)
+        # Lower thresholds for medical images to ensure detections are not filtered out
+        results = run_sam3_inference(pil_image, prompt_text, threshold=0.1, mask_threshold=0.0)
         
         final_mask = None
         if results and 'masks' in results and results['masks'] is not None: