Spaces:
Running on Zero
Running on Zero
Download app.py from hugging-apps/p2echo: direct link, hf CLI and curl.
- Browser
- Download file 8.4 kB
-
https://huggingface.co/spaces/hugging-apps/p2echo/resolve/main/app.py
- Command line
-
hf download hf://spaces/hugging-apps/p2echo/app.py
-
curl -L -o app.py https://huggingface.co/spaces/hugging-apps/p2echo/resolve/main/app.py
8.4 kB
| """P2Echo: text-prompted multi-view echocardiography segmentation demo. | |
| Upload an echocardiography frame, pick the view and the cardiac structures to | |
| segment, and P2Echo returns a colour-coded segmentation mask following the | |
| structured text prompt built from your selection. | |
| """ | |
| import os | |
| os.environ.setdefault("PYTORCH_CUDA_ALLOC_CONF", "expandable_segments:True") | |
| import spaces # noqa: E402 (must precede torch) | |
| import torch # noqa: E402 | |
| import gradio as gr # noqa: E402 | |
| import numpy as np # noqa: E402 | |
| from PIL import Image # noqa: E402 | |
| from p2echo import P2Echo, TextEncoder # noqa: E402 | |
| from p2echo.data import normalize_image # noqa: E402 | |
| from p2echo.prompts import VIEW_STRUCTURES, structure_prompt # noqa: E402 | |
| DEVICE = "cuda" | |
| IMAGE_SIZE = 256 | |
| TEXT_MODEL = "Qwen/Qwen3-Embedding-0.6B" | |
| CHECKPOINT = "p2echo.pth" # shipped in the Space repo | |
| # ---------------------------------------------------------------- model load | |
| text_encoder = TextEncoder(TEXT_MODEL).to(DEVICE) | |
| model = P2Echo(text_dim=text_encoder.dim).to(DEVICE).eval() | |
| state = torch.load(CHECKPOINT, map_location="cpu", weights_only=True)["model"] | |
| model.load_state_dict(state) | |
| print("P2Echo loaded:", sum(p.numel() for p in model.parameters()) / 1e6, "M params") | |
| STRUCTURE_COLORS = { | |
| "LV": (255, 60, 60), | |
| "MYO": (60, 200, 255), | |
| "LA": (255, 220, 60), | |
| "RV": (80, 130, 255), | |
| "RA": (220, 80, 255), | |
| } | |
| def preprocess(pil_image): | |
| """PIL -> normalized [1,3,256,256] tensor, exactly like p2echo.data.EchoDataset.""" | |
| img = pil_image.convert("RGB").resize((IMAGE_SIZE, IMAGE_SIZE), Image.BILINEAR) | |
| arr = np.array(img).transpose(2, 0, 1) | |
| tensor = torch.from_numpy(arr).float() / 255.0 | |
| return normalize_image(tensor).unsqueeze(0) | |
| def _segment_impl(image, view, structures, threshold): | |
| """Reference implementation returning (overlay PIL, gallery list) as plain CPU objects.""" | |
| view = view.upper() | |
| structures = [s for s in structures if s] or ["LV"] | |
| requested = [s for s in structures if s in VIEW_STRUCTURES[view]] | |
| if not requested: | |
| raise gr.Error(f"None of {structures} is visualized in the {view} view. " | |
| f"Choose one of: {', '.join(VIEW_STRUCTURES[view])}.") | |
| prompts = [structure_prompt(view, s) for s in requested] | |
| emb = text_encoder(prompts, DEVICE).unsqueeze(0) # [1, N, dim] | |
| img_t = preprocess(image).to(DEVICE) | |
| with torch.autocast("cuda", dtype=torch.bfloat16): | |
| logits = model(img_t, emb)[0].float() # [1, N, H, W] | |
| probs = torch.sigmoid(logits)[0].cpu().numpy() # [N, H, W] | |
| n = len(requested) | |
| base = image.convert("RGB").resize((IMAGE_SIZE, IMAGE_SIZE), Image.BILINEAR) | |
| overlay = np.array(base, dtype=np.float32) | |
| for i, s in enumerate(requested): | |
| m = probs[i] >= threshold | |
| color = np.array(STRUCTURE_COLORS[s], dtype=np.float32) | |
| overlay[m] = overlay[m] * 0.45 + color * 0.55 | |
| overlay_img = Image.fromarray(overlay.clip(0, 255).astype(np.uint8)) | |
| gallery = [(base, "input"), (overlay_img, "overlay")] | |
| for i, s in enumerate(requested): | |
| m = probs[i] >= threshold | |
| color = np.array(STRUCTURE_COLORS[s], dtype=np.float32) | |
| tile = np.full((IMAGE_SIZE, IMAGE_SIZE, 3), 255, dtype=np.float32) | |
| tile[m] = color | |
| gallery.append((Image.fromarray(tile.clip(0, 255).astype(np.uint8)), f"{s} ({float(probs[i].max()):.2f} max)")) | |
| return overlay_img, gallery | |
| def segment(image, view, structures=("LV",), threshold=0.5): | |
| """Segment the selected cardiac structures in an echocardiography frame. | |
| Args: | |
| image: input echocardiography image (any size; resized to 256x256). | |
| view: echocardiographic view, one of 2CH / 3CH / 4CH / PSAX. | |
| structures: list of structures to segment (LV, MYO, LA, RV, RA). | |
| threshold: probability threshold binarising each structure's mask. | |
| Returns: | |
| (overlay image, gallery of input/overlay/per-structure colour masks). | |
| """ | |
| if image is None: | |
| raise gr.Error("Please upload an echocardiography image first.") | |
| return _segment_impl(image, view, structures, float(threshold)) | |
| VIEW_DESCRIPTIONS = { | |
| "4CH": "Apical 4-Chamber", | |
| "2CH": "Apical 2-Chamber", | |
| "3CH": "Apical 3-Chamber", | |
| "PSAX": "Parasternal Short-Axis", | |
| } | |
| def update_structures(view): | |
| """Restrict the structure checkboxes to those visible in the chosen view.""" | |
| return gr.update(choices=list(VIEW_STRUCTURES[view.upper()]), | |
| value=["LV"] if "LV" in VIEW_STRUCTURES[view.upper()] | |
| else list(VIEW_STRUCTURES[view.upper()])[:1]) | |
| CSS = """ | |
| #col-container { max-width: 1210px; margin: 0 auto; } | |
| .dark .gradio-container { color: var(--body-text-color); } | |
| """ | |
| with gr.Blocks(title="P2Echo · Text-Prompted Echo Segmentation") as demo: | |
| with gr.Column(elem_id="col-container"): | |
| gr.Markdown( | |
| """ | |
| # 🫀 P2Echo · Text-Prompted Echocardiography Segmentation | |
| **P2Echo** segments cardiac structures (LV, MYO, LA, RV, RA) across four echo views | |
| (2CH, 3CH, 4CH, PSAX) with a single 37.5M-parameter model controlled by a structured | |
| text prompt naming the view and the requested structure. | |
| Upload a frame, pick the view, select the structures — the demo builds the paper's | |
| view-aware prompt (e.g. *"Apical 4-Chamber (4CH): … Segment the LV: The LV appears as | |
| a bullet-shaped chamber forming the apex."*) and runs one forward pass. | |
| [Paper](https://arxiv.org/abs/2611.13425) · [Code](https://github.com/moeinheidari7829/P2Echo) · | |
| [Weights](https://huggingface.co/moein99/P2Echo) — research use only, not for clinical decisions. | |
| """ | |
| ) | |
| with gr.Row(): | |
| with gr.Column(scale=5): | |
| input_img = gr.Image(type="pil", label="Echocardiography frame", height=320) | |
| view = gr.Radio(list(VIEW_DESCRIPTIONS.keys()), value="4CH", label="View", | |
| info="Standard echocardiographic view of the frame") | |
| structures = gr.CheckboxGroup(list(VIEW_STRUCTURES["4CH"]), value=["LV"], | |
| label="Structures to segment", | |
| info="Restricted to structures visible in the view") | |
| threshold = gr.Slider(0.1, 0.9, value=0.5, step=0.05, label="Threshold") | |
| run_btn = gr.Button("Segment", variant="primary") | |
| prompt_preview = gr.Textbox(label="Text prompt sent to the model", lines=4, interactive=False) | |
| with gr.Column(scale=7): | |
| overlay_out = gr.Image(type="pil", label="Prediction overlay", height=320) | |
| gallery_out = gr.Gallery(label="Input · Overlay · Per-structure masks", | |
| columns=4, height=300) | |
| gr.Examples( | |
| examples=[ | |
| ["examples/4ch.png", "4CH"], | |
| ["examples/2ch.png", "2CH"], | |
| ["examples/psax.png", "PSAX"], | |
| ], | |
| inputs=[input_img, view], | |
| outputs=[overlay_out, gallery_out], | |
| fn=segment, | |
| cache_examples=True, | |
| cache_mode="lazy", | |
| label="Examples (click one, then Segment)", | |
| ) | |
| gr.Markdown( | |
| """**Reference**: Heidari et al., *Prompt2Echo (P2Echo): Text-Prompted Multi-View | |
| Echocardiography Segmentation with One Model*, arXiv:2611.13425. Text encoder: | |
| frozen Qwen3-Embedding-0.6B.""" | |
| ) | |
| view.change(update_structures, inputs=view, outputs=structures) | |
| def _preview(view, structures): | |
| chosen = [s for s in structures if s in VIEW_STRUCTURES[view.upper()]][:1] | |
| if not chosen: | |
| return gr.update() | |
| return structure_prompt(view.upper(), chosen[0]) | |
| structures.change(_preview, inputs=[view, structures], outputs=prompt_preview) | |
| view.change(_preview, inputs=[view, structures], outputs=prompt_preview) | |
| run_btn.click( | |
| segment, | |
| inputs=[input_img, view, structures, threshold], | |
| outputs=[overlay_out, gallery_out], | |
| api_name="segment", | |
| ) | |
| demo.launch(theme=gr.themes.Citrus(), css=CSS, mcp_server=True) | |