import streamlit as st
from transformers import pipeline
from PIL import Image
import numpy as np
import cv2

st.set_page_config(page_title="Détection de fractures osseuses")
st.title("Détection de fractures osseuses par rayons X")

@st.cache_resource
def load_model():
    return pipeline(
        "object-detection",
        model="anirban22/detr-resnet-50-med_fracture",
        threshold=0.1
    )

model = load_model()

uploaded_file = st.file_uploader("Téléchargez une image radiographique", type=["jpg", "jpeg", "png"])

if uploaded_file:
    # Convert uploaded file to PIL Image
    image = Image.open(uploaded_file).convert('RGB')
    if image.size[0] > 800:
        ratio = 800.0 / image.size[0]
        size = (800, int(image.size[1] * ratio))
        image = image.resize(size, Image.Resampling.LANCZOS)
    
    # Pass PIL Image directly to model
    predictions = model(image)
    
    col1, col2 = st.columns(2)
    
    with col1:
        st.image(image, caption="Image originale", use_container_width=True)
    
    with col2:
        img_array = np.array(image)
        for pred in predictions:
            box = pred['box']
            score = pred['score']
            x1, y1, x2, y2 = [int(i) for i in [box['xmin'], box['ymin'], box['xmax'], box['ymax']]]
            cv2.rectangle(img_array, (x1, y1), (x2, y2), (255, 0, 0), 3)
            text = f"Fracture: {score:.2f}"
            cv2.putText(img_array, text, (x1, y1-10), cv2.FONT_HERSHEY_SIMPLEX, 0.7, (255, 0, 0), 2)
        
        st.image(img_array, caption="Détection des fractures", use_container_width=True)
    
    if predictions:
        st.subheader(f"🚨 {len(predictions)} fractures détectées:")
        for idx, pred in enumerate(predictions, 1):
            st.warning(f"Fracture {idx}: Confiance {pred['score']*100:.1f}%")
    else:
        st.warning("⚠️ Aucune fracture détectée")