Medical Research Platform for Pathology and Disease Identification

Artificial intelligence (AI) in healthcare has progressed from early rule-based expert systems to modern deep learning methods particularly convolutional neural networks (CNNs) that now achieve strong performance in medical image interpretation. Despite this progress, clinical adoption remains limited by interconnected barriers: diagnostic variability and error under high workload, a global shortage of imaging specialists, fragmented clinical data across incompatible systems, and the “black-box” nature of many models, which reduces transparency and clinician trust. Additional concerns include limited generalization to diverse populations and the practical burden of secure deployment, training, and compliance.

This thesis proposes an AI-driven Medical Research Platform for Pathology and Disease Identification, delivered as a cloud-based minimum viable product (MVP) that integrates validated pretrained models via transfer learning as a proof of concept. The initial implementation targets two high-impact imaging domains: thoracic chest X-rays (e.g., cardiomegaly and effusion) and ophthalmic/retinal imaging (including conditions such as cataracts, diabetic retinopathy, and glaucoma). To improve interpretability and clinical usability, the platform incorporates explainable AI techniques (e.g., Grad-CAM) and provides structured evaluation outputs for continuous quality assurance.

A central contribution is the platform’s role-based, collaborative ecosystem spanning four user groups. Researchers can validate and assess models using curated datasets and standardized metrics (accuracy, specificity, F1-score, confusion matrices) supported by visual analytics. Clinics manage medical teams and monitor AI-assisted outcomes transparently, while doctors the primary decision-makers upload images, review AI-generated preliminary analyses, annotate or adjust findings, and retain full responsibility for communicating final diagnoses to patients. Patients, invited by doctors, can remotely upload images, track their diagnostic history, and view only physician-approved results, preserving clinical integrity. The system also integrates the ChatGPT API to generate detailed, understandable narratives for doctors and patients by combining model outputs with relevant case context while excluding personally identifiable information.

The scope emphasizes secure access (authentication and role-based control), workflow fit, and ongoing doctor-led model validation (including approval gates and continuous feedback). The project does not aim to create novel model architectures or achieve medical-device regulatory certification in this phase; future work includes expanding to additional modalities (CT, MRI, ultrasound), on-premises deployment options, and formal clinical validation pathways.