Diagnosense

Abstract

This project presents development and design of an intelligent web based diagnosis platform for anatomical support and classification of lung disease. The model uses hybrid deep learning structure to extract local features from PET/CT images using ResNet50V2 followed by global contextual information extracted from PET/CT images by a Vision Transformer (ViTAttention) module. Two-stage training with preliminary feature extraction and fine-tuning was applied in order to increase the accuracy and generalisation of the model. The model was well-trained and tested on the LUNG-PET-CT-DX dataset, with a good demonstration on clinically significant measures such as the accuracy, precision, recall, F1-score, and AUC. With regard to the support of interpretability, LIME-based visualizations were included, which gave an explanation of how the model makes decisions. In order to support the system, a domain-specific lung anatomy chatbot was also added to the system, on a Retrieval-Augmented Generation (RAG) basis. The chatbot answers the queries of the users by searching the relevant content from a PDF based knowledge base through Facebook AI Similarity Search (FAISS) vector search and provides medically backed responses with the help of the Gemma API. The platform is delivered through a responsive web interface built with the MERN stack and FastAPI that allows for real-time image classification and interactive anatomical guidance. The system is intended at radiologists, clinicians, and medical researchers as a precise and interpretable mechanism that is accessible to allow early detection of the lung disease and anatomical learning.