Multi-Modal Diagnostic Assistant

Using CNN as a tool for clinical dermantology cases.

Project Overview

Summary

This project introduces a prototype Multi-Modal Diagnostic Assistant for veterinary dermatology, blending machine learning with an expert logic layer. A Convolutional Neural Network trained on skin lesion images provides baseline predictions, which are then refined by a clinical rules engine incorporating symptoms and patient history. This hybrid approach effectively emulates the nuanced diagnostic process of a veterinarian without requiring a massive multi-modal dataset.

I am developing a prototype “Multi-Modal Diagnostic Assistant” for veterinary dermatology. This tool aims to go beyond simple image recognition by emulating a veterinarian’s diagnostic process. It will analyze skin lesion images alongside essential clinical context (symptoms, patient history) to deliver intelligent and nuanced assessments.

Methodology (Our Approach)

To achieve this with available resources, I will employ a hybrid methodology combining machine learning with an expert-driven rules engine:

Machine Learning Core

  • Convolutional Neural Network (CNN): Train a CNN in Google Colab using a public dataset of dog skin images to serve as the “vision expert,” providing baseline predictions based on visual data.

Expert Logic Layer

  • Differential Diagnosis Engine: Build within the Streamlit application, this engine will use rules and logic based on clinical veterinary knowledge to refine diagnoses.

Intelligent Fusion

  • Process: The app first retrieves the image-based prediction from the AI model. It then incorporates user-entered clinical data (e.g., itch level, lesion location) into the expert logic to refine the prediction, add context, and provide clinically relevant explanations.
  • Advantage: This approach bypasses the need for a comprehensive multi-modal dataset while creating a prototype that functions as a sophisticated, expert-guided system.

Project Phases

Phase 1 (This Weekend’s Goal - MVP)

  • Train the base image model in Google Colab.
  • Develop and deploy a functional Streamlit web app with:
    • UI for clinical data entry.
    • Core rule-based logic for at least 4 dermatological conditions.
  • Deliverable: A live, shareable web link for the app.

Phase 2 (Future Refinement)

  • Enhance the image model with additional data.
  • Expand the logic engine to include more conditions and complex diagnostic rules.
  • Improve the user interface and experience.

Phase 3 (Future Expansion)

  • Explore developing a true multi-modal neural network if a suitable dataset is found or created.
  • Integrate Explainable AI (XAI) techniques to visually highlight areas of focus in images.

Why & How This Project is Useful

This project stands out as an exceptional portfolio piece for the following reasons:

Demonstrates Advanced Thinking

  • Showcases creative problem-solving by addressing the challenge of missing datasets with a practical, intelligent solution.

Showcases a Hybrid Skillset

  • Combines technical expertise (Python, TensorFlow, Streamlit) with domain knowledge in veterinary science and medicine, a rare and valuable skill combination.

Tells a Compelling Story

  • This is not just another image classifier; it’s a narrative of bridging technology and real-world clinical applications, highlighting a deep understanding of the problem and a unique ability to address it effectively.

Multi-modal diagnostic tool

After using MobileNetV2


© Balaji Ramanathan