Update app.py

app.py

@@ -18,6 +18,9 @@ device = torch.device("cpu")
 save_dir = "/home/user/app/saved_predictions"
 os.makedirs(save_dir, exist_ok=True)
 
+# Placeholder image for invalid / non-fundus uploads
+invalid_img = Image.new("RGB", (224, 224), color=(200, 200, 200))
+
 # -----------------------
 # Load model
 # -----------------------
@@ -45,28 +48,26 @@ transform = transforms.Compose([
 def looks_like_fundus(image):
     """
     Very lightweight heuristic to guess if an image looks like a retinal fundus scan.
-
-    This is NOT a medical-grade classifier – it is only used to show a warning
-    if the image is very unlikely to be a retina.
+    NOT a medical-grade classifier – only used to suppress predictions on
+    obviously wrong images.
     """
     img = np.array(image.convert("L").resize((224, 224)))
 
-    # Central square (potential retina) vs border
+    # Central square vs border
     center = img[40:184, 40:184]
 
     border_mask = np.ones_like(img, dtype=bool)
     border_mask[40:184, 40:184] = False
     border_pixels = img[border_mask]
 
-    # Safety fallback
     if border_pixels.size == 0:
         return True
 
     center_mean = center.mean()
     border_mean = border_pixels.mean()
 
-    border_dark_ratio = np.mean(border_pixels < 40)   # dark background
-    center_bright_ratio = np.mean(center > 80)        # bright retina
+    border_dark_ratio = np.mean(border_pixels < 40)
+    center_bright_ratio = np.mean(center > 80)
 
     cond_contrast = center_mean - border_mean > 15
     cond_border_dark = border_dark_ratio > 0.3
@@ -79,16 +80,18 @@ def looks_like_fundus(image):
 # Predict and save
 # -----------------------
 def predict_retinopathy(image):
-    # 1. Soft validation (no blocking – just warning text)
-    if looks_like_fundus(image):
-        warning = ""
-    else:
-        warning = (
-            "⚠️ The uploaded image may not be a retinal fundus scan. "
-            "This system is intended for use with ophthalmic retinal images only.\n\n"
+    is_fundus = looks_like_fundus(image)
+
+    # If it's clearly not a fundus image: DO NOT give a DR/NoDR decision
+    if not is_fundus:
+        return (
+            invalid_img,
+            "⚠️ Image not recognised as a retinal fundus scan.\n\n"
+            "No Diabetic Retinopathy assessment has been generated. "
+            "Please upload a valid ophthalmic retinal fundus image."
         )
 
-    # 2. Normal pipeline
+    # Normal pipeline for likely fundus images
     timestamp = datetime.datetime.now().strftime("%Y%m%d_%H%M%S")
     img = image.convert("RGB").resize((224, 224))
     img_tensor = transform(img).unsqueeze(0).to(device)
@@ -115,9 +118,7 @@ def predict_retinopathy(image):
     filename = f"{timestamp}_{label}_{confidence:.2f}.png"
     cam_pil.save(os.path.join(save_dir, filename))
 
-    prediction_text = f"{label} (Confidence: {confidence:.2f})"
-
-    return cam_pil, warning + prediction_text
+    return cam_pil, f"{label} (Confidence: {confidence:.2f})"
 
 
 # -----------------------
@@ -127,7 +128,7 @@ demo = gr.Interface(
     fn=predict_retinopathy,
     inputs=gr.Image(type="pil", label="Upload Retinal Image"),
     outputs=[
-        gr.Image(type="pil", label="Grad-CAM Heatmap"),
+        gr.Image(type="pil", label="Grad-CAM / Status"),
         gr.Text(label="Prediction")
     ],
     title="OpthaDetect – AI Retinal Screening",
@@ -138,7 +139,7 @@ demo = gr.Interface(
     article=(
         "⚕️ **OpthaDetect** is an AI-powered ophthalmic decision-support tool. "
         "It highlights retinal risk regions using Grad-CAM for better clinical interpretability. "
-        "This tool does not replace clinical judgement and should be used alongside professional assessment."
+        "Outputs should always be reviewed alongside clinical judgement."
     )
 )