main.py

import base64
import io
import json
import numpy as np
import torch
from fastapi import FastAPI, HTTPException
from huggingface_hub import hf_hub_download
from PIL import Image
from pydantic import BaseModel
from torchvision import transforms

import timm
import segmentation_models_pytorch as smp
import albumentations as A
from albumentations.pytorch import ToTensorV2
import os

# --- 1. SETUP: Create the FastAPI app ---
app = FastAPI(title="AI Skin Lesion Analyzer API")

# --- Global variables ---
DEVICE = "cpu"
segmentation_model = None
classification_model = None
knowledge_base = None
idx_to_class_abbr = {0: 'MEL', 1: 'NV', 2: 'BCC', 3: 'AKIEC', 4: 'BKL', 5: 'DF', 6: 'VASC'}
transform_segment = None
transform_classify = None

class ImageRequest(BaseModel):
    image_base64: str

@app.on_event("startup")
def load_assets():
    global segmentation_model, classification_model, knowledge_base, transform_segment, transform_classify

    # Haddii hore loo load gareeyay, iska dhaaf
    if segmentation_model is not None and classification_model is not None and knowledge_base is not None:
        print("🔁 Models and knowledge base already loaded. Skipping reloading.")
        return

    print("--> API starting up: This may take a few minutes...")

    # Use /tmp for writable cache directory
    cache_dir = "/tmp/models_cache"
    os.makedirs(cache_dir, exist_ok=True)

    # Load Segmentation Model
    try:
        print("    Downloading UNet segmentation model...")
        seg_model_path = hf_hub_download(
            repo_id="sheikh987/unet-isic2018",
            filename="unet_full_data_best_model.pth",
            cache_dir=cache_dir
        )
        segmentation_model = smp.Unet("resnet34", encoder_weights=None, in_channels=3, classes=1).to(DEVICE)
        segmentation_model.load_state_dict(torch.load(seg_model_path, map_location=DEVICE))
        segmentation_model.eval()
        print("    ✅ Segmentation model loaded.")
    except Exception as e:
        print(f"!!! FATAL: Could not load segmentation model: {e}")

    # Load Classification Model
    try:
        print("    Downloading EfficientNet classification model...")
        class_model_path = hf_hub_download(
            repo_id="sheikh987/efficientnet-isic",
            filename="efficientnet_augmented_best.pth",
            cache_dir=cache_dir
        )
        classification_model = timm.create_model('efficientnet_b3', pretrained=False, num_classes=7).to(DEVICE)
        classification_model.load_state_dict(torch.load(class_model_path, map_location=DEVICE))
        classification_model.eval()
        print("    ✅ Classification model loaded.")
    except Exception as e:
        print(f"!!! FATAL: Could not load classification model: {e}")

    # Load Knowledge Base
    try:
        with open('knowledge_base.json', 'r') as f:
            knowledge_base = json.load(f)
        print("    ✅ Knowledge base loaded.")
    except Exception as e:
        print(f"!!! FATAL: Could not load knowledge_base.json: {e}")

    # Define Image Transforms
    transform_segment = A.Compose([
        A.Resize(256, 256),
        A.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225], max_pixel_value=255.0),
        ToTensorV2()
    ])

    transform_classify = transforms.Compose([
        transforms.Resize((300, 300)),
        transforms.ToTensor(),
        transforms.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225])
    ])

    print("\n--> API is ready to accept requests.")

@app.post("/analyze/")
async def analyze_image(request: ImageRequest):
    if not all([segmentation_model, classification_model, knowledge_base]):
        raise HTTPException(status_code=503, detail="Models not loaded yet. Please retry shortly.")

    try:
        image_data = base64.b64decode(request.image_base64)
        image = Image.open(io.BytesIO(image_data)).convert("RGB")
        image_np = np.array(image)
    except Exception:
        raise HTTPException(status_code=400, detail="Invalid base64 image data provided.")

    # Stage 1: Segmentation
    augmented_seg = transform_segment(image=image_np)
    seg_input_tensor = augmented_seg['image'].unsqueeze(0).to(DEVICE)
    with torch.no_grad():
        seg_logits = segmentation_model(seg_input_tensor)
    seg_mask = (torch.sigmoid(seg_logits) > 0.5).float().squeeze().cpu().numpy()

    if seg_mask.sum() < 200:
        return {"status": "Failed", "message": "No lesion could be clearly identified in the image."}

    # Stage 2: Crop and classify
    rows, cols = np.any(seg_mask, axis=1), np.any(seg_mask, axis=0)
    rmin, rmax = np.where(rows)[0][[0, -1]]
    cmin, cmax = np.where(cols)[0][[0, -1]]
    padding = 15
    rmin, rmax = max(0, rmin - padding), min(image_np.shape[0], rmax + padding)
    cmin, cmax = max(0, cmin - padding), min(image_np.shape[1], cmax + padding)
    cropped_image_pil = Image.fromarray(image_np[rmin:rmax, cmin:cmax])

    class_input_tensor = transform_classify(cropped_image_pil).unsqueeze(0).to(DEVICE)
    with torch.no_grad():
        class_logits = classification_model(class_input_tensor)
        probabilities = torch.nn.functional.softmax(class_logits, dim=1)

    confidence, predicted_idx = torch.max(probabilities, 1)
    confidence_percent = confidence.item() * 100

    if confidence_percent < 50.0:
        return {"status": "Inconclusive", "message": f"Model confidence ({confidence_percent:.2f}%) is below the 75% threshold."}

    # Stage 3: Return result
    predicted_abbr = idx_to_class_abbr[predicted_idx.item()]
    info = knowledge_base.get(predicted_abbr, {})

    return {
        "status": "Success",
        "prediction": info,
        "abbreviation": predicted_abbr,
        "confidence": f"{confidence_percent:.2f}%"
    }

@app.get("/")
def root():
    return {"message": "AI Skin Lesion Analyzer API is running."}