Our Team

Our core team synergizes deep clinical dermatology expertise (MD), foundational AI/data engineering (final-year Info. Eng. student), & robust mathematical skills (EE with control sys. math depth). We are planning on adding a Biomedical Prof. and/or IE Prof. for senior research oversight. Crucially, funds will secure a dedicated AI Research Engineer to lead advanced AI dev, ensuring a potent blend of domain knowledge, technical skill, & specialized research capability for this project.

Project details

The Challenge: Critical Gaps in Melanoma Diagnosis and AI Fairness

Melanoma, the most aggressive form of skin cancer, poses a significant global health challenge. Early and accurate diagnosis is paramount for improving patient outcomes, yet current diagnostic pathways face limitations. Visual assessment can be subjective, and while AI has shown promise, many existing models function as "black boxes," lacking the transparency crucial for clinical trust and adoption. Furthermore, a critical issue is algorithmic bias: AI models trained on unrepresentative datasets often underperform on darker skin tones, exacerbating health disparities. This RFP's call for neural-symbolic DNNs is timely, as these architectures offer a powerful avenue to address these shortcomings.

Our Vision: EXAI-Melanoma – An Explainable & Equitable Neural-Symbolic Solution

This project, "EXAI-Melanoma", proposes to develop and demonstrate a novel neural-symbolic system for the early, explainable, and equitable diagnosis of melanoma. We aim to integrate the strengths of cutting-edge neural networks for interpretable feature extraction with the power of symbolic AI for incorporating clinical knowledge and ensuring logical reasoning. Our system is designed to not only improve diagnostic accuracy but also to provide transparent justifications for its predictions and to perform robustly and fairly across diverse patient populations.

Core Technological Pillars

Our solution rests on two primary neural-symbolic technologies, with potential for a third experiential learning component:

3.1. Kolmogorov-Arnold Networks (KANs) for Interpretable Image Analysis & Equity

Traditional CNNs, while powerful, are often opaque. KANs, inspired by the Kolmogorov-Arnold representation theorem, offer a promising alternative. Instead of fixed activation functions within nodes, KANs feature learnable activation functions on edges. This intrinsic property allows for greater interpretability, as these learned functions (e.g., splines) can be visualized and analyzed to understand how input features are transformed.

• Application: We will explore KANs (or KAN- GNN hybrids like KA-GNNs if deemed optimal during research) to extract features from dermoscopic images.
• Explainability: The goal is to identify and visualize image features critical for diagnosis directly from the KAN's structure, moving beyond post-hoc explanations.
• Equity Focus: A core objective is to train and evaluate our KAN-based models on diverse datasets, explicitly including a wide range of skin tones (e.g., leveraging datasets like Fitzpatrick17k, or augmenting ISIC/HAM10000 with strategies to improve representation). We will investigate how KANs' functional flexibility might help in capturing relevant features across varied skin presentations more effectively than traditional models and implement fairness-aware learning techniques.

3.2. PyNeuraLogic for Higher-Order Symbolic Reasoning & Data Integration:

PyNeuraLogic, a framework combining differentiable logic programming with deep learning (especially GNNs), is ideal for embedding symbolic knowledge and performing rule-based reasoning.

1. Clinical Guideline Embedding: We will encode established dermatological guidelines (e.g., ABCD-E criteria, 7-point checklist) and risk factors as differentiable logic rules within PyNeuraLogic.
2. Patient Context Integration: We plan to use GNNs within the PyNeuraLogic framework to model and integrate structured patient data (e.g., Fitzpatrick skin type, age, gender, family history, lesion history, potentially from EMR-like structures). This allows for a holistic diagnostic assessment.
3. Medical Ontology Grounding: SNOMED CT will be explored to structure the symbolic knowledge, ensuring consistency and interoperability for lesion types, features, and patient conditions. This directly addresses the RFP's interest in structured learning.

3.3. (Optional Exploration) AIRIS-Inspired Experiential Rule Refinement:

We may explore an AIRIS-like experiential learning component. This could involve a simulated environment where the system learns to refine or discover new diagnostic rules based on interactions with curated case studies and feedback. This aligns with the RFP's interest in experiential learning for rule generation.

Integration and MeTTa Proof of Concept (POC)

The KAN-based image analysis module will provide interpretable image features and initial classifications. These outputs, along with patient contextual data, will feed into the PyNeuraLogic system. PyNeuraLogic will then apply its embedded symbolic rules and learned GNN representations to derive a final, reasoned diagnostic suggestion, along with an explanation trace.

A key deliverable will be a Proof of Concept (POC) implemented in the MeTTa language. This POC will demonstrate the core functionalities:

1. Ingestion of (simulated or real, anonymized) image features and patient data.
2. Application of the KAN-inspired feature analysis and PyNeuraLogic rule-based reasoning.
3. Generation of an explainable diagnostic output.

The MeTTa POC will showcase how these neural-symbolic components can be integrated within the Hyperon AGI framework, aligning with SingularityNET's vision.

Alignment with RFP "Neural-symbolic DNN architectures"

This project directly addresses the RFP's main purpose and key questions:

1. Embedding Logic Rules: We embed both expert-defined clinical guidelines and rules potentially refined via experiential learning.
2. Improving Reasoning in DNNs: PyNeuraLogic, working with GNNs and KAN outputs, enhances reasoning beyond pure data-driven approaches.
3. Human Interpretability & Explainability: A core design goal, achieved through KANs' intrinsic properties and PyNeuraLogic's symbolic nature.
4. Learning from Small Data: Symbolic knowledge helps constrain models and improve generalization, crucial for medical domains often facing data scarcity for specific subgroups.
5. Bridging Data-Driven and Symbolic Reasoning: This is the essence of our neural-symbolic architecture.
6. Structured Learning: Addressed by using SNOMED CT and GNNs for structured patient data.
7. Complex System Dynamics: The interaction between image features, patient history, and clinical rules within a dynamic diagnostic process will be explored.

Innovation and Key Differentiators

1. Novel application of KANs for interpretable and equitable dermoscopic image analysis.
2. A tightly coupled neural-symbolic architecture specifically designed for melanoma diagnosis, integrating image, patient, and clinical rule data.
3. A strong focus on algorithmic fairness and explainability as primary design objectives.
4. Commitment to a MeTTa POC, demonstrating real-world utility within the SingularityNET ecosystem.

Expected Outcomes & Deliverables (as per RFP)

1. Comprehensive Survey and Evaluation: A detailed analysis of KANs and PyNeuraLogic (and related NeSy architectures) for this dermatological application, including their strengths and limitations.
2. Architectural Comparisons: Evaluation of how the proposed architecture performs in embedding rules and reasoning, addressing scalability and flexibility.
3. Proof of Concept (POC): A functional POC in MeTTa demonstrating the core system.
4. System Dynamics Demonstration: Evidence that the approach can drive interesting diagnostic dynamics, improving reasoning.
5. Explainability Report: Detailing how the architecture enhances human interpretability.
6. Small Data Learning Analysis: Comparison of learning with and without symbolic knowledge.
7. Multiparadigmality Showcase: Demonstrating the bridge between data-driven and symbolic reasoning.
8. Structured Learning Application: Showcasing the use of medical ontologies.
9. Final Report, Code (open-sourced), Framework Demonstration, Documentation.

Plan for Fairness and Ethical AI

We are committed to responsible AI development. Our strategy includes:

1. Prioritizing diverse datasets for training and testing, with a focus on varied skin tones.
2. Implementing bias detection and mitigation techniques during model development.
3. Ensuring transparency in model predictions through our XAI focus.
4. Regular consultation with our dermatologist MD to ensure clinical relevance and safety.
5. Adherence to ethical AI principles for healthcare.

High-Level Project Plan & Team Strengths

This 6-month project will follow agile principles, broadly encompassing:

1. Milestone 1 (Research & Design): Literature review, detailed KAN/PyNeuraLogic architecture design, data acquisition/preparation strategy, MeTTa integration plan, detailed research plan.
2. Milestone 2 (Development & Initial Testing): Implementation of KAN image module, PyNeuraLogic rule/GNN module, initial integration, preliminary testing on benchmark datasets.
3. Milestone 3 (POC, Evaluation & Reporting): MeTTa POC development, comprehensive evaluation (accuracy, fairness, explainability), final report, code documentation, and dissemination.

Our core team provides a unique blend of direct clinical dermatology expertise (MD), foundational AI/data engineering skills (final-year Information Engineering student with a strong math foundation from an Electrical Engineering collaborator), and we are in advanced discussions to bring on a Biomedical Professor and/or Information Engineering Professor for senior research leadership. A key use of the requested funds ($95,000) will be to onboard an experienced AI/ML Research Engineer specializing in neural-symbolic systems and/or MeTTa to accelerate advanced development and ensure the project's ambitious technical goals are met. This strategic team augmentation will create a powerhouse capable of delivering impactful results.

Impact and Contribution

Successful completion will:

• Advance the field of explainable and equitable AI in medical diagnosis.
• Provide a novel neural-symbolic framework for melanoma detection.
• Contribute a valuable use-case and POC to the SingularityNET/MeTTa ecosystem.
• Offer insights into applying KANs and PyNeuraLogic to complex, real-world problems.
• Ultimately, hold the potential to improve clinical decision support and reduce disparities in melanoma outcomes.

Open Source Licensing

Links and references

Team Members: 

1. Giga Hidjrika Aura Adkhy, Information Engineering - Giga Hidjrika Aura Adkhy | LinkedIn 
2. Eta Auria Latiefa, Medical Doctor, Dermatovenereologist Resident -  Eta Auria Latiefa | LinkedIn
3. Azfar Azdi Arfakhsyad, Electrical Engineer, Control Systems Expert -  Azfar Azdi Arfakhsyad | LinkedIn

Pending: 

1. Ridwan Wicaksono, ST, M.Eng., Ph.D., Assistant Professor in Biomedics - ACADSTAFF UGM
2. Ahmad Ataka Awwalur Rizqi, S.T., Ph.D., Robotics Researcher - Academic Staff