app.py

import streamlit as st
import google.generativeai as genai
import os
from PIL import Image
import io # Needed for handling image bytes
from typing import Optional, Tuple, Any # For type hinting

# --- Page Configuration (MUST BE THE FIRST STREAMLIT COMMAND) ---
st.set_page_config(
    page_title="AI Clinical Support Demonstrator",
    layout="wide",
    initial_sidebar_state="expanded"
)

# --- Introductory Explanation ---
st.markdown(
    """
    ### Welcome to the AI Clinical Support Demonstrator
    This application demonstrates how Generative AI (Google Gemini) can be guided to assist with analyzing clinical information.
    - **Agentic Text Analysis:** Simulates a structured reasoning process on clinical text to identify key findings, suggest considerations, and note information gaps.
    - **Medical Image Analysis:** Provides descriptive observations of potential anomalies in medical images.

    **Crucially, this tool is for demonstration purposes ONLY. It does NOT provide medical advice or diagnosis.**
    """
)
st.markdown("---") # Visual separator

# --- Configuration and Initialization ---

# Securely load API key (Secrets > Env Var)
GEMINI_API_KEY = st.secrets.get("GEMINI_API_KEY", os.environ.get("GEMINI_API_KEY"))

# Configure Gemini Client
genai_client_configured = False
if GEMINI_API_KEY:
    try:
        genai.configure(api_key=GEMINI_API_KEY)
        genai_client_configured = True
    except Exception as e:
        st.error(f"Fatal Error: Failed to configure Google Generative AI. Check API Key. Details: {e}", icon="🚨")
        st.stop()
else:
    st.error("⚠️ Gemini API Key not found. Please configure `GEMINI_API_KEY` in Streamlit secrets or environment variables.", icon="🔑")
    st.stop()

# Initialize models
TEXT_MODEL_NAME = 'gemini-1.5-pro-latest' # For agentic text reasoning
VISION_MODEL_NAME = 'gemini-1.5-flash'   # For image analysis

# Use session state to initialize models only once
if 'models_initialized' not in st.session_state:
    st.session_state.models_initialized = False
    st.session_state.text_model = None
    st.session_state.vision_model = None

if genai_client_configured and not st.session_state.models_initialized:
    try:
        st.session_state.text_model = genai.GenerativeModel(TEXT_MODEL_NAME)
        st.session_state.vision_model = genai.GenerativeModel(VISION_MODEL_NAME)
        st.session_state.models_initialized = True
        # Display success message subtly in sidebar perhaps, or remove if too noisy
        # st.sidebar.success("AI Models Ready.", icon="✅")
    except Exception as e:
        st.error(f"Fatal Error: Failed to initialize Gemini models. Text: {TEXT_MODEL_NAME}, Vision: {VISION_MODEL_NAME}. Details: {e}", icon="💥")
        st.stop()
elif not genai_client_configured:
    # Should have stopped already, but defensive check
    st.error("AI Models could not be initialized due to configuration issues.", icon="🚫")
    st.stop()


# --- Core AI Interaction Functions ---

# Refined AGENTIC prompt for Text Analysis with formatting request
AGENTIC_TEXT_ANALYSIS_PROMPT_TEMPLATE = """
**Simulated Clinical Reasoning Agent Task:**

**Role:** AI assistant simulating an agentic clinical reasoning process to support a healthcare professional by structuring information, generating possibilities, and suggesting investigation pathways based *strictly* on the provided text. **This is NOT a diagnosis.**

**Input Data:** Unstructured clinical information (e.g., symptoms, history, basic findings).

**Output Format:** Please structure your response using Markdown headings for each step (e.g., `## 1. Information Extraction`).

**Simulated Agentic Steps (Perform sequentially):**

1.  **Information Extraction & Structuring:**
    *   Key Demographics (Age, Sex if provided).
    *   Primary Symptoms/Signs.
    *   Relevant Medical History.
    *   Pertinent Negatives (if mentioned).

2.  **Differential Considerations Generation:**
    *   Based *only* on Step 1, list **potential differential considerations** (possible conditions).
    *   **Use cautious language:** "could be consistent with," "warrants consideration," "less likely but possible." **AVOID definitive statements.**
    *   Briefly justify each consideration based on findings.

3.  **Information Gap Analysis:**
    *   Identify critical missing information (e.g., lab results, imaging, exam specifics, duration/onset).

4.  **Suggested Next Steps for Investigation (for Clinician):**
    *   Propose logical next steps a **healthcare professional might consider**.
    *   Categorize (e.g., Further History, Exam Points, Labs, Imaging).
    *   Frame as *suggestions* (e.g., "Consider ordering...", "Assessment of X may be informative").

5.  **Mandatory Disclaimer:** Conclude with: "This AI-generated analysis is for informational support only. It is **NOT** a diagnosis and cannot replace the judgment of a qualified healthcare professional."

**Input Clinical Information:**
---
{text_input}
---

**Agentic Analysis:**
"""

# Refined prompt for Image Analysis with formatting request
IMAGE_ANALYSIS_PROMPT_TEMPLATE = """
**Medical Image Analysis Request:**

**Context:** Analyze the provided medical image. User may provide additional context or questions.

**Output Format:** Please structure your response using Markdown headings for each step (e.g., `## 1. Visible Structures`).

**Task:**

1.  **Describe Visible Structures:** Briefly describe main anatomical structures.
2.  **Identify Potential Anomalies:** Point out areas that *appear* abnormal or deviate from typical presentation (e.g., "potential opacity," "altered signal intensity," "possible asymmetry"). Use cautious, descriptive language. **This is not a definitive finding.**
3.  **Correlate with User Prompt (if provided):** Address specific user questions based *only* on visual information. State if the image cannot answer the question.
4.  **Limitations:** State that image quality, view, lack of clinical context, and the AI's nature limit analysis.
5.  **Mandatory Disclaimer:** Explicitly state this is AI visual analysis, not radiological interpretation or diagnosis, requiring review by a qualified professional with clinical context.

**User's Additional Context/Question (if any):**
---
{user_prompt}
---

**Image Analysis:**
"""

def run_agentic_text_analysis(text_input: str) -> Tuple[Optional[str], Optional[str]]:
    """Sends clinical text to the configured text model for simulated agentic analysis."""
    if not text_input or not text_input.strip():
        return None, "Input text cannot be empty."
    if not st.session_state.models_initialized or not st.session_state.text_model:
         return None, "Text analysis model not initialized."

    try:
        prompt = AGENTIC_TEXT_ANALYSIS_PROMPT_TEMPLATE.format(text_input=text_input)
        response = st.session_state.text_model.generate_content(prompt)

        if response.parts:
            return response.text, None
        elif response.prompt_feedback.block_reason:
             return None, f"Analysis blocked by safety filters: {response.prompt_feedback.block_reason.name}. Review input."
        else:
             candidate = response.candidates[0] if response.candidates else None
             if candidate and candidate.finish_reason != "STOP":
                 return None, f"Analysis stopped prematurely. Reason: {candidate.finish_reason.name}."
             else:
                 return None, "Received an empty or unexpected response from the AI model."

    except Exception as e:
        print(f"ERROR in run_agentic_text_analysis: {e}") # Basic server-side log
        st.error("An error occurred during text analysis.", icon="🚨") # User-facing error
        return None, "An internal error occurred during text analysis. Please try again later."

def analyze_medical_image(image_file: Any, user_prompt: str = "") -> Tuple[Optional[str], Optional[str]]:
    """Sends a medical image to the configured vision model for analysis."""
    if not image_file:
        return None, "Image file cannot be empty."
    if not st.session_state.models_initialized or not st.session_state.vision_model:
         return None, "Image analysis model not initialized."

    try:
        try:
            image = Image.open(image_file)
            # Convert to RGB ensure compatibility, common requirement for vision models
            if image.mode != 'RGB':
                 image = image.convert('RGB')
        except Exception as img_e:
             return None, f"Error opening or processing the uploaded image file: {img_e}"

        prompt_text = IMAGE_ANALYSIS_PROMPT_TEMPLATE.format(user_prompt=user_prompt if user_prompt else "N/A")
        model_input = [prompt_text, image]
        response = st.session_state.vision_model.generate_content(model_input)

        if response.parts:
            return response.text, None
        elif response.prompt_feedback.block_reason:
            return None, f"Image analysis blocked by safety filters: {response.prompt_feedback.block_reason.name}. This may relate to sensitive content policies."
        else:
             candidate = response.candidates[0] if response.candidates else None
             if candidate and candidate.finish_reason != "STOP":
                 return None, f"Image analysis stopped prematurely. Reason: {candidate.finish_reason.name}."
             else:
                 return None, "Received an empty or unexpected response from the AI model for image analysis."

    except Exception as e:
        print(f"ERROR in analyze_medical_image: {e}") # Basic server-side log
        st.error("An error occurred during image analysis.", icon="🖼️") # User-facing error
        return None, "An internal error occurred during image analysis. Please try again later."


# --- Streamlit User Interface ---

def main():
    # Page title and model info
    st.title("🤖 AI Clinical Support Demonstrator")
    st.caption(f"Utilizing: Text Model ({TEXT_MODEL_NAME}), Vision Model ({VISION_MODEL_NAME})")

    # --- CRITICAL DISCLAIMER ---
    # Positioned prominently after title/intro
    st.warning(
        """
        **🔴 IMPORTANT SAFETY & USE DISCLAIMER 🔴**
        *   This tool **DEMONSTRATES** AI capabilities. It **DOES NOT** provide medical advice or diagnosis.
        *   **Agentic Text Analysis:** Simulates reasoning on text input. Output is illustrative, not diagnostic.
        *   **Image Analysis:** Provides observations on images. Output is **NOT** a radiological interpretation.
        *   AI analysis lacks full clinical context, may be inaccurate, and **CANNOT** replace professional judgment.
        *   **ALWAYS consult qualified healthcare professionals** for diagnosis and treatment.
        *   **PRIVACY:** Do **NOT** upload identifiable patient information (PHI) without explicit consent and adherence to all privacy laws (e.g., HIPAA). You are responsible for the data you input.
        """,
        icon="⚠️"
    )
    st.markdown("---")

    # --- Sidebar Controls ---
    st.sidebar.header("Analysis Options")
    input_method = st.sidebar.radio(
        "Select Analysis Type:",
        ("Agentic Text Analysis", "Medical Image Analysis"),
        key="input_method_radio",
        help="Choose 'Agentic Text Analysis' for reasoning simulation on clinical text, or 'Medical Image Analysis' for visual observations on images."
    )
    st.sidebar.markdown("---") # Visual separator

    # --- Main Area Layout (Input and Output Columns) ---
    col1, col2 = st.columns(2)

    analysis_result = None # Initialize results variables for this run
    error_message = None
    output_header = "Analysis Results" # Default header

    # --- Column 1: Input Area ---
    with col1:
        st.header("Input Data")

        # Conditional Input UI based on selection
        if input_method == "Agentic Text Analysis":
            st.subheader("Clinical Text for Agentic Analysis")
            st.caption("Please ensure data is de-identified before pasting.")
            text_input = st.text_area(
                "Paste clinical information:",
                height=350,
                placeholder="Example: 68yo male, sudden SOB & pleuritic chest pain post-flight. HR 110, SpO2 92% RA. No known cardiac hx...",
                key="text_input_area"
            )
            analyze_button_key = "analyze_text_button"
            analyze_button_label = "▶️ Run Agentic Text Analysis"

            if st.button(analyze_button_label, key=analyze_button_key, type="primary"):
                if text_input:
                    with st.spinner("🧠 Simulating agentic reasoning..."):
                        analysis_result, error_message = run_agentic_text_analysis(text_input)
                    output_header = "Simulated Agentic Analysis Output"
                else:
                    st.warning("Please enter clinical text to analyze.", icon="☝️")

        elif input_method == "Medical Image Analysis":
            st.subheader("Medical Image for Analysis")
            st.caption("Upload a de-identified medical image (PNG, JPG, JPEG).")
            image_file = st.file_uploader(
                "Choose an image file:",
                type=["png", "jpg", "jpeg"],
                key="image_uploader"
            )
            user_image_prompt = st.text_input(
                "Optional: Add context or specific question for image analysis:",
                placeholder="Example: 'Describe findings in the lung fields' or 'Any visible fractures?'",
                key="image_prompt_input"
            )
            analyze_button_key = "analyze_image_button"
            analyze_button_label = "🖼️ Analyze Medical Image"

            if image_file:
                # Display preview immediately after upload inside the input column
                st.image(image_file, caption="Uploaded Image Preview", use_column_width=True)

            if st.button(analyze_button_label, key=analyze_button_key, type="primary"):
                if image_file:
                    with st.spinner("👁️ Analyzing image..."):
                        analysis_result, error_message = analyze_medical_image(image_file, user_image_prompt)
                    output_header = "Medical Image Analysis Output"
                else:
                    st.warning("Please upload an image file to analyze.", icon="☝️")

    # --- Column 2: Output Area ---
    with col2:
        st.header(output_header)

        # Display results or errors if an analysis was attempted
        # Check if button was pressed and corresponding result/error exists
        button_pressed = st.session_state.get(analyze_button_key, False) if 'analyze_button_key' in locals() else False

        if button_pressed and (analysis_result or error_message):
             if error_message:
                 st.error(f"Analysis Failed: {error_message}", icon="❌")
             elif analysis_result:
                 # Use st.markdown to render potential formatting from AI
                 st.markdown(analysis_result)
        elif not button_pressed :
            st.info("Analysis results will appear here after providing input and clicking the corresponding analysis button.")
        # Add a condition for button pressed but no input provided (already handled by st.warning in col1)


    # --- Sidebar Explanations ---
    st.sidebar.markdown("---")
    st.sidebar.header("About The Prompts")
    with st.sidebar.expander("View Agentic Text Prompt Structure", icon="📄"):
        st.markdown(f"```plaintext\n{AGENTIC_TEXT_ANALYSIS_PROMPT_TEMPLATE.split('---')[0]} ... [Input Text] ...\n```")
        st.caption("Guides the AI through structured reasoning steps for text.")
    with st.sidebar.expander("View Image Analysis Prompt Structure", icon="🖼️"):
        st.markdown(f"```plaintext\n{IMAGE_ANALYSIS_PROMPT_TEMPLATE.split('---')[0]} ... [User Prompt] ...\n```")
        st.caption("Guides the AI to describe visual features and potential anomalies in images.")

    st.sidebar.markdown("---")
    st.sidebar.error(
        "**Ethical Use Reminder:** AI in medicine requires extreme caution. This tool is for demonstration and education, not clinical practice. Verify all information and rely on professional expertise.",
        icon = "⚕️"
        )

# --- Main Execution Guard ---
if __name__ == "__main__":
    # Check if models are initialized before running the main UI components
    if st.session_state.models_initialized:
        main()
    else:
        # If models failed to initialize, errors would have been shown already.
        # We might add a placeholder or further message here if needed.
        st.info("Waiting for model initialization...") # Or check specific errors if needed
        # The script might stop earlier if initialization fails critically.