app.py

import gradio as gr
import os
import tensorflow as tf
import numpy as np
import requests
import time

from langchain_groq import ChatGroq
from langchain.agents import initialize_agent
from langchain.prompts import PromptTemplate
from langchain_core.runnables import RunnableSequence
from langchain.tools import StructuredTool


IMG_HEIGHT = 256
IMG_WIDTH = 256


model_path = "unet_model.h5"
if not os.path.exists(model_path):
    hf_url = "https://huggingface.co/rishirajbal/UNET_plus_plus_Brain_segmentation/resolve/main/unet_model.h5"
    print(f"Downloading model from {hf_url}...")
    with requests.get(hf_url, stream=True) as r:
        r.raise_for_status()
        with open(model_path, "wb") as f:
            for chunk in r.iter_content(chunk_size=8192):
                f.write(chunk)

print("Loading model...")
model = tf.keras.models.load_model(model_path, compile=False)



def classify_image_and_stats(image_input):
    img = tf.image.resize(image_input, [IMG_HEIGHT, IMG_WIDTH])
    img_norm = img / 255.0
    img_batch = np.expand_dims(img_norm, axis=0)

    prediction = model.predict(img_batch)[0]
    mask = (prediction > 0.5).astype(np.uint8)

    if mask.ndim == 3 and mask.shape[-1] == 1:
        mask = np.squeeze(mask, axis=-1)

    tumor_area = np.sum(mask)
    total_area = IMG_HEIGHT * IMG_WIDTH
    tumor_ratio = tumor_area / total_area

    tumor_label = "Tumor Detected" if tumor_ratio > 0.00385 else "No Tumor Detected"

    overlay = np.array(img)
    red_mask = np.zeros_like(overlay)
    red_mask[..., 0] = mask * 255

    overlay_img = np.clip(0.6 * overlay + 0.4 * red_mask, 0, 255).astype(np.uint8)

    stats = {
        "tumor_area": int(tumor_area),
        "total_area": total_area,
        "tumor_ratio": tumor_ratio,
        "tumor_label": tumor_label
    }

    return overlay_img, stats



def rishigpt_handler(image_input, groq_api_key):
    os.environ["GROQ_API_KEY"] = groq_api_key

    overlay_img, stats = classify_image_and_stats(image_input)

    def segment_brain_tool(input_text: str) -> str:
        return (
            f"Tumor label: {stats['tumor_label']}. "
            f"Tumor area: {stats['tumor_area']}. "
            f"Ratio: {stats['tumor_ratio']:.4f}."
        )

    tool = StructuredTool.from_function(
        segment_brain_tool,
        name="segment_brain",
        description="Provide tumor segmentation stats for the MRI image."
    )

    llm = ChatGroq(
        model="meta-llama/llama-4-scout-17b-16e-instruct",
        temperature=0.4
    )

    agent = initialize_agent(
        tools=[tool],
        llm=llm,
        agent="zero-shot-react-description",
        verbose=False
    )

    user_query = "Give me the segmentation details"
    classification = agent.run(user_query)

    prompt = PromptTemplate(
        input_variables=["result"],
        template=(
            "You are a compassionate AI radiologist. "
            "Read this tumor analysis result: {result}. "
            "Summarize the situation for the patient in natural paragraphs, calm, clear tone, with next steps."
        )
    )

    chain = prompt | llm
    final_text = chain.invoke({"result": classification}).content.strip()


    displayed_text = ""
    for char in final_text:
        displayed_text += char
        time.sleep(0.015)
        yield overlay_img, displayed_text


inputs = [
    gr.Image(type="numpy", label="Upload Brain MRI Slice"),
    gr.Textbox(type="password", label="Groq API Key")
]

outputs = [
    gr.Image(type="numpy", label="Overlay: Brain MRI + Tumor Mask"),
    gr.Textbox(label="Doctor's Explanation")
]

if __name__ == "__main__":
    gr.Interface(
        fn=rishigpt_handler,
        inputs=inputs,
        outputs=outputs,
        title="RishiGPT Medical Brain Segmentation",
        description="UNet++ Brain Tumor Segmentation with live mask overlay, detailed stats, and human-like typing explanation."
    ).launch()