Towards explainable oral cancer recognition: Screening on imperfect images via Informed Deep Learning and Case-Based...

Area of Science:

• Medical Informatics
• Artificial Intelligence in Healthcare
• Oncology Diagnostics

Background:

• Oral squamous cell carcinoma (OSCC) diagnosis is often delayed, leading to poor outcomes and high costs.
• Current AI diagnostic tools lack transparency, hindering clinical adoption.
• Existing explainable AI (XAI) methods prioritize developer needs over clinical user requirements for insights.

Purpose of the Study:

• To develop a cost-efficient, transparent, and accurate computerized system for early OSCC detection.
• To integrate medical knowledge into deep learning (DL) models for improved clinical relevance.
• To enhance XAI by providing clinically meaningful explanations for AI-driven diagnoses.

Main Methods:

• Proposed a novel approach combining Case-Based Reasoning (CBR) for visual explanations and Informed Deep Learning (IDL) to integrate medical knowledge.
• Utilized an ensemble architecture within the DL workflow to robustly handle imperfect data, including labeling inaccuracies and artifacts.
• Conducted experimental benchmarks on a dataset from collaborative medical centers.

Main Results:

• The IDL approach achieved an accuracy of 85%, significantly outperforming standard DL methods (77% accuracy).
• IDL demonstrated superior human-centered explainability, generating insights more aligned with clinical user demands compared to traditional XAI.
• The system effectively handled data imperfections, showcasing its robustness in real-world clinical scenarios.

Conclusions:

• IDL offers a promising solution for accurate and transparent early detection of oral squamous cell carcinoma.
• Integrating medical knowledge and employing CBR for explanations enhances AI system utility in clinical practice.
• The developed system addresses key challenges in OSCC screening, including cost, accuracy, and clinical interpretability.