GemmaSight

GemmaSight is an explainable AI-powered pathology assistant designed to support the early detection of cancer from digital pathology slide images. Traditional cancer diagnosis often requires specialized laboratories, expert pathologists, and significant processing time, making timely diagnosis difficult, especially in remote and underserved regions. GemmaSight addresses these challenges by providing fast, transparent, and AI-assisted preliminary screening. The system first processes a pathology slide image using Google's Pathology Foundation Model, which extracts rich visual embeddings that capture microscopic tissue features. These embeddings are then analyzed by a custom-trained classifier to predict whether the tissue is Cancer or Non-Cancer. Our classifier achieved approximately 94% accuracy with an AUC score of 0.98, demonstrating strong predictive performance. To improve trust and interpretability, GemmaSight integrates a FAISS-based similarity retrieval system. Instead of providing only a prediction, the system retrieves the three most similar pathology cases from its embedding database, allowing clinicians to compare the current sample with previously analyzed cases. Additionally, an occlusion heatmap highlights the image regions that most influenced the model's prediction, providing visual evidence of the AI's decision-making process. Finally, the pathology image, prediction, retrieved similar cases, and explainability outputs are passed to the MedGemma Vision-Language Model, which generates a concise, high-level medical report in approximately 30 seconds. GemmaSight is designed to assist—not replace—medical professionals. By combining accurate classification, explainable AI, similarity retrieval, and automated report generation into a single workflow, it enables faster, more transparent, and more accessible AI-assisted pathology, helping clinicians make informed decisions and improving access to quality healthcare in resource-limited settings.