Upload folder using huggingface_hub

.gitattributes

@@ -33,3 +33,8 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 *.zip filter=lfs diff=lfs merge=lfs -text
 *.zst filter=lfs diff=lfs merge=lfs -text
 *tfevents* filter=lfs diff=lfs merge=lfs -text
+examples/real_colon_004-001_frame13.jpg filter=lfs diff=lfs merge=lfs -text
+examples/real_colon_004-001_frame25.jpg filter=lfs diff=lfs merge=lfs -text
+examples/real_colon_004-001_frame37.jpg filter=lfs diff=lfs merge=lfs -text
+examples/real_colon_004-001_frame49.jpg filter=lfs diff=lfs merge=lfs -text
+examples/real_colon_004-001_frame61.jpg filter=lfs diff=lfs merge=lfs -text

README.md

@@ -1,13 +1,51 @@
 ---
-title: True Colon Polyp Detection
-emoji: ⚡
-colorFrom: indigo
-colorTo: pink
+title: TRUE-Colon Polyp Detection
+emoji: 🩺
+colorFrom: purple
+colorTo: yellow
 sdk: gradio
 sdk_version: 6.24.0
-python_version: '3.12'
 app_file: app.py
-pinned: false
+short_description: RT-DETR polyp detector on REAL-Colon colonoscopy frames
+python_version: "3.12"
+startup_duration_timeout: 30m
 ---
 
-Check out the configuration reference at https://huggingface.co/docs/hub/spaces-config-reference
+# TRUE-Colon: Real-Time Polyp Detection
+
+Interactive demo of the **RT-DETR** detector from the paper *TRUE-Colon: Exposing a
+Consistent Transfer Asymmetry in Real-Time Polyp Detection* (MICCAI 2026 EndoLINA
+Workshop). The model was trained on **REAL-Colon** — 60 complete, unedited colonoscopy
+procedures — and detects polyps (lesions) in individual colonoscopy frames.
+
+## Usage
+
+Upload a colonoscopy frame, adjust the confidence and IoU thresholds, and click
+**Detect Polyps**. The model draws green bounding boxes around detected lesions with
+confidence scores.
+
+> ⚠️ **Research demo only — NOT for clinical use.** The model has not been validated
+> prospectively or cleared by any regulator.
+
+## Model
+
+- **Checkpoint**: [`sdoerrich97/true_colon_rtdetr_realcolon_s0`](https://huggingface.co/sdoerrich97/true_colon_rtdetr_realcolon_s0)
+- **Architecture**: RT-DETR (via Ultralytics)
+- **Input**: 640 × 640
+- **Classes**: 1 (`lesion`)
+- **License**: AGPL-3.0 (inherited from Ultralytics training)
+
+## Example images
+
+Examples are from the [REAL-Colon](https://doi.org/10.1038/s41597-024-03359-0) dataset
+(Biffi et al., *Scientific Data* 2024, CC BY 4.0) and [CVC-ClinicDB](https://polyp.grand-challenge.org/CVCClinicDB/) (CC BY 4.0).
+
+## Citation
+
+```bibtex
+@article{doerrich2026truecolon,
+  title={TRUE-Colon: Exposing a Consistent Transfer Asymmetry in Real-Time Polyp Detection},
+  author={Sebastian Doerrich and Andreas Franz Schwab and Francesco {Di Salvo} and Shyam Nandan Rai and Hanh Huyen My Nguyen and Christian Ledig},
+  year={2026}, eprint={2608.13711}, archivePrefix={arXiv}, primaryClass={eess.IV}
+}
+```

app.py

@@ -0,0 +1,205 @@
+"""TRUE-Colon: Real-Time Polyp Detection demo.
+
+Loads the RT-DETR checkpoint from the TRUE-Colon paper (MICCAI 2026 EndoLINA Workshop)
+and runs inference on colonoscopy frames, drawing bounding boxes around detected polyps.
+
+Research demo only — NOT for clinical use.
+"""
+
+import spaces  # MUST be first
+import os
+import cv2
+import numpy as np
+import torch
+from huggingface_hub import hf_hub_download
+from ultralytics import YOLO
+import gradio as gr
+
+MODEL_ID = "sdoerrich97/true_colon_rtdetr_realcolon_s0"
+CLASS_NAME = "lesion"
+# Green-ish box color (BGR for cv2)
+BOX_COLOR = (0, 255, 0)
+
+# Download and load model at module scope
+_weights_path = hf_hub_download(MODEL_ID, "model.pt")
+model = YOLO(_weights_path)
+
+
+def draw_detections(image: np.ndarray, results, conf_threshold: float) -> np.ndarray:
+    """Draw bounding boxes on the image from Ultralytics results.
+
+    Args:
+        image: Input image as numpy array (RGB).
+        results: Ultralytics prediction results.
+        conf_threshold: Confidence threshold for display.
+
+    Returns:
+        Annotated image as numpy array (RGB).
+    """
+    annotated = image.copy()
+    h, w = annotated.shape[:2]
+
+    for result in results:
+        boxes = result.boxes
+        for box in boxes:
+            conf = float(box.conf[0])
+            if conf < conf_threshold:
+                continue
+            cls = int(box.cls[0])
+            x1, y1, x2, y2 = box.xyxy[0].cpu().numpy()
+            x1, y1, x2, y2 = int(x1), int(y1), int(x2), int(y2)
+
+            # Draw box
+            cv2.rectangle(annotated, (x1, y1), (x2, y2), BOX_COLOR, 3)
+
+            # Draw label background
+            label = f"{CLASS_NAME} {conf:.2f}"
+            (label_w, label_h), _ = cv2.getTextSize(
+                label, cv2.FONT_HERSHEY_SIMPLEX, 0.7, 2
+            )
+            cv2.rectangle(
+                annotated,
+                (x1, y1 - label_h - 10),
+                (x1 + label_w, y1),
+                BOX_COLOR,
+                -1,
+            )
+            cv2.putText(
+                annotated,
+                label,
+                (x1, y1 - 5),
+                cv2.FONT_HERSHEY_SIMPLEX,
+                0.7,
+                (0, 0, 0),
+                2,
+                cv2.LINE_AA,
+            )
+
+    return annotated
+
+
+@spaces.GPU(duration=30)
+def detect(
+    image: np.ndarray,
+    conf_threshold: float = 0.30,
+    iou_threshold: float = 0.50,
+) -> np.ndarray:
+    """Detect polyps in a colonoscopy frame.
+
+    Runs the TRUE-Colon RT-DETR detector on the input image and returns an
+    annotated copy with bounding boxes around detected lesions.
+
+    Args:
+        image: Colonoscopy frame as an image.
+        conf_threshold: Minimum detection confidence to display.
+        iou_threshold: NMS IoU threshold.
+
+    Returns:
+        Annotated image with detection boxes drawn.
+    """
+    if image is None:
+        return None
+
+    # Ultralytics expects RGB; Gradio passes RGB
+    results = model.predict(
+        source=image,
+        conf=conf_threshold,
+        iou=iou_threshold,
+        imgsz=640,
+        verbose=False,
+    )
+
+    annotated = draw_detections(image, results, conf_threshold)
+    return annotated
+
+
+CSS = """
+#col-container { max-width: 1100px; margin: 0 auto; }
+.dark .gradio-container { color: var(--body-text-color); }
+"""
+
+EXAMPLES = [
+    ["examples/real_colon_004-001_frame13.jpg"],
+    ["examples/real_colon_004-001_frame25.jpg"],
+    ["examples/real_colon_004-001_frame37.jpg"],
+    ["examples/real_colon_004-001_frame49.jpg"],
+    ["examples/cvc_2.png"],
+    ["examples/cvc_100.png"],
+]
+
+with gr.Blocks(theme=gr.themes.Citrus(), css=CSS) as demo:
+    with gr.Column(elem_id="col-container"):
+        gr.Markdown(
+            """
+            # 🩺 TRUE-Colon: Real-Time Polyp Detection
+            **RT-DETR** trained on REAL-Colon (60 full colonoscopy procedures) for polyp detection.
+            From the paper *TRUE-Colon: Exposing a Consistent Transfer Asymmetry in Real-Time Polyp Detection* (MICCAI 2026 EndoLINA).
+
+            [📄 Paper](https://arxiv.org/abs/2608.13711) | [🐍 Code](https://github.com/sdoerrich97/true-colon) | [⚖️ Weights](https://huggingface.co/sdoerrich97/true_colon_rtdetr_realcolon_s0)
+
+            > ⚠️ **Research demo only — NOT for clinical use.** The model has not been validated prospectively or cleared by any regulator.
+            """
+        )
+
+        with gr.Row():
+            with gr.Column(scale=1):
+                input_image = gr.Image(
+                    label="Colonoscopy Frame",
+                    type="numpy",
+                    sources=["upload", "clipboard"],
+                )
+                conf_slider = gr.Slider(
+                    label="Confidence Threshold",
+                    minimum=0.05,
+                    maximum=0.95,
+                    step=0.05,
+                    value=0.30,
+                )
+                iou_slider = gr.Slider(
+                    label="NMS IoU Threshold",
+                    minimum=0.10,
+                    maximum=0.95,
+                    step=0.05,
+                    value=0.50,
+                )
+                run_btn = gr.Button("Detect Polyps", variant="primary")
+
+            with gr.Column(scale=1):
+                output_image = gr.Image(
+                    label="Detection Result",
+                    type="numpy",
+                )
+
+        run_btn.click(
+            fn=detect,
+            inputs=[input_image, conf_slider, iou_slider],
+            outputs=[output_image],
+            api_name="detect",
+        )
+
+        gr.Examples(
+            examples=EXAMPLES,
+            inputs=[input_image],
+            outputs=[output_image],
+            fn=detect,
+            cache_examples=True,
+            cache_mode="lazy",
+        )
+
+        gr.Markdown(
+            """
+            ### Example image sources
+            Examples from the [REAL-Colon](https://doi.org/10.1038/s41597-024-03359-0) dataset (Biffi et al., *Scientific Data* 2024, CC BY 4.0) and [CVC-ClinicDB](https://polyp.grand-challenge.org/CVCClinicDB/) (CC BY 4.0).
+
+            ### Citation
+            ```bibtex
+            @article{doerrich2026truecolon,
+              title={TRUE-Colon: Exposing a Consistent Transfer Asymmetry in Real-Time Polyp Detection},
+              author={Sebastian Doerrich and Andreas Franz Schwab and Francesco {Di Salvo} and Shyam Nandan Rai and Hanh Huyen My Nguyen and Christian Ledig},
+              year={2026}, eprint={2608.13711}, archivePrefix={arXiv}, primaryClass={eess.IV}
+            }
+            ```
+            """
+        )
+
+demo.launch(mcp_server=True)

requirements.txt

@@ -0,0 +1,3 @@
+ultralytics>=8.3.223
+opencv-python-headless
+numpy