app.py

import os
import h5py
import numpy as np
import gradio as gr
import plotly.graph_objects as go
from railnet_model import RailNetSystem

from huggingface_hub import hf_hub_download

os.environ['KMP_DUPLICATE_LIB_OK'] = 'True'
os.environ["CUDA_VISIBLE_DEVICES"] = "0"

model = RailNetSystem.from_pretrained("Tournesol-Saturday/railNet-tooth-segmentation-in-CBCT-image").cuda()

model.load_weights(from_hub=True, repo_id="Tournesol-Saturday/railNet-tooth-segmentation-in-CBCT-image")

def render_plotly_volume(pred, x_eye=1.25, y_eye=1.25, z_eye=1.25):
    downsample_factor = 2
    pred_ds = pred[::downsample_factor, ::downsample_factor, ::downsample_factor]

    fig = go.Figure(data=go.Volume(
        x=np.repeat(np.arange(pred_ds.shape[0]), pred_ds.shape[1] * pred_ds.shape[2]),
        y=np.tile(np.repeat(np.arange(pred_ds.shape[1]), pred_ds.shape[2]), pred_ds.shape[0]),
        z=np.tile(np.arange(pred_ds.shape[2]), pred_ds.shape[0] * pred_ds.shape[1]),
        value=pred_ds.flatten(),
        isomin=0.5,
        isomax=1.0,
        opacity=0.1,
        surface_count=1,
        colorscale=[[0, 'rgb(255, 0, 0)'], [1, 'rgb(255, 0, 0)']],
        showscale=False
    ))

    fig.update_layout(
        scene=dict(
            xaxis=dict(visible=False),
            yaxis=dict(visible=False),
            zaxis=dict(visible=False),
            camera=dict(eye=dict(x=x_eye, y=y_eye, z=z_eye))
        ),
        margin=dict(l=0, r=0, b=0, t=0)
    )
    return fig

def handle_example(filename):
    repo_id = "Tournesol-Saturday/railNet-tooth-segmentation-in-CBCT-image"
    h5_path = hf_hub_download(repo_id=repo_id, filename=f"example_input_file/{filename}")
   
    with h5py.File(h5_path, "r") as f:
        image = f["image"][:]
        label = f["label"][:]
    
    name = filename.replace(".h5", "")
    pred, dice, jc, hd, asd = model(image, label, "./output", name)
    
    fig = render_plotly_volume(pred)
    
    img_path = f"./output/{name}_img.nii.gz"
    pred_path = f"./output/{name}_pred.nii.gz"
    
    metrics = f"Dice: {dice:.4f}, Jaccard: {jc:.4f}, 95HD: {hd:.2f}, ASD: {asd:.2f}"
    
    return metrics, pred, fig, img_path, pred_path

def clear_all():
    return "", None, None, None, None

with gr.Blocks() as demo:
    gr.HTML("<div style='text-align: center; font-size: 22px; font-weight: bold;'>🦷 Demo of RailNet: A CBCT Tooth Segmentation System</div>")
    gr.HTML("<div style='text-align: center; font-size: 15px'>✅ Steps: Select a CBCT example file (.h5)  →  Automatic inference and metrics display  →  View 3D segmentation result (Mouse drag and scroll wheel zooming)</div>")

    gr.HTML("""
    <style>
    .code-style {
        font-family: monospace;
        background-color: #2f363d;
        color: #ffffff;
        padding: 2px 6px;
        border-radius: 4px;
        font-size: 90%;
    }
    </style>

    <div style='font-size: 15px; font-weight: bold;'>
        📂 Step 1: Select a <span class='code-style'>.h5</span> example file from the <span class='code-style'>example_input_file</span> folder in our
        <a href='https://huggingface.co/Tournesol-Saturday/railNet-tooth-segmentation-in-CBCT-image' target='_blank' style='text-decoration: none; color: #1f6feb; font-weight: bold;'>
        Hugging Face model
        </a> repository. 
    </div>
    """)

    example_files = ["CBCT_01.h5", "CBCT_02.h5", "CBCT_03.h5", "CBCT_04.h5"] 
    dropdown = gr.Dropdown(choices=example_files, label="Example File", value=example_files[0])


    with gr.Row():
        clear_btn = gr.Button("清除", variant="secondary")
        submit_btn = gr.Button("提交", variant="primary")

    gr.HTML("<div style='font-size: 15px; font-weight: bold;'>📊 Step 2: Metrics (Dice, Jaccard, 95HD, ASD)</div>")
    result_text = gr.Textbox()
    hidden_pred = gr.State(value=None) 

    gr.HTML("<div style='font-size: 15px; font-weight: bold;'>👁️ Step 3: 3D Visualisation</div>")
    plot_output = gr.Plot()

    gr.HTML("<div style='font-size: 15px; font-weight: bold;'>⬇️ Step 4: Download <span class='code-style'>NIfTI</span> files for accurate 1:1 visualization using <span class='code-style'>ITK-SNAP</span> software</div>")
    with gr.Row():
        hidden_img_file = gr.File(label="Download Original Image", interactive=False)
        hidden_pred_file = gr.File(label="Download Segmentation Result", interactive=False)

    submit_btn.click(
        fn=handle_example,
        inputs=[dropdown],
        outputs=[result_text, hidden_pred, plot_output, hidden_img_file, hidden_pred_file]
    )

    clear_btn.click(
        fn=clear_all,
        inputs=[],
        outputs=[result_text, hidden_pred, plot_output, hidden_img_file, hidden_pred_file]
    )

demo.launch()