SALLY FAPOHUNDS

My name is Sally Fapohunds, and I am focused on developing AI-assisted diagnostic systems for gastrointestinal diseases, with a particular emphasis on chronic atrophic gastritis (CAG) detection and risk stratification under white-light endoscopy. With a background in medical image computing and clinical informatics, I am motivated by the need to bridge diagnostic variability and enhance early detection accuracy. CAG is a known precancerous condition, yet its subtle endoscopic features make it difficult for clinicians to detect consistently. My goal is to leverage artificial intelligence to improve visual interpretation and ensure timely, precise diagnosis.

My work centers on building deep learning models that can analyze white-light endoscopic images and identify CAG-related features such as mucosal thinning, loss of gastric folds, and glandular atrophy. I employ convolutional neural networks, attention-based visual explainability modules, and multi-task learning to jointly perform lesion detection and risk classification. To support real-world deployment, I integrate data from multicenter image repositories and ensure the model is trained on diverse populations. The system is further validated with expert annotations and clinical outcomes to align with diagnostic gold standards.

One of the key challenges in AI for endoscopy is clinical trust and interpretability. To address this, I embed explainable AI mechanisms into the diagnostic workflow—generating heatmaps, visual highlights, and confidence scores that assist gastroenterologists in understanding and validating AI outputs. The system is designed not to replace the clinician but to assist in improving detection sensitivity and reducing inter-observer variability. My ultimate aim is to build an AI platform that integrates seamlessly into hospital systems and supports personalized surveillance strategies for patients with chronic gastritis.

As a researcher at the intersection of AI and gastroenterology, my mission is to develop clinically robust, ethically sound diagnostic tools that enhance decision-making without adding complexity to clinical workflows. I plan to expand this research to incorporate longitudinal patient data for prognosis prediction and to explore multimodal fusion with pathology and serological biomarkers. Looking forward, I hope to collaborate with international research hospitals and regulatory bodies to help define AI safety and performance standards for digestive disease diagnosis. Through translational research, I aim to contribute to earlier detection, better outcomes, and equitable care in gastrointestinal health.

Medical Terminology & Multimodal Fusion: CAG diagnosis requires specialized endoscopic, pathological, and clinical language. GPT-3.5 lacks pretraining on GI and pathology corpora, limiting its ability to accurately integrate multimodal features.

Chain-of-Thought Reasoning: The model must produce coherent causal chains linking pathology, clinical data, and predictions. GPT-3.5 often exhibits logical jumps or misses key steps in long-context reasoning.

Dual Task of Classification & Regression: Simultaneous staging classification and risk-probability regression necessitate multi-objective optimization—GPT-3.5’s context window and capacity are insufficient for stable, high-precision performance across both tasks.

Injection of Domain Knowledge: We need to embed OLGA/OLGIM ontologies, up-to-date clinical guidelines, and regional epidemiology into the model weights. GPT-3.5 fine-tuning struggles with catastrophic forgetting under such constraints.

GPT-4’s larger parameter scale and extended context window, combined with staged fine-tuning, can deliver high accuracy alongside stable chain-of-thought explanations and robust multimodal integration, reducing hallucinations by ≥40%—making it the only viable path for a professional-grade AI-assisted diagnostic system.