Embedded Files
Early and accurate detection of colon cancer is essential for improving patient outcomes, yet the limited interpretability of deep learning models often reduces clinicians' trust in AI-assisted diagnosis. This paper presents RAG-ViT, an intelligent framework that integrates a Vision Transformer (ViT) with Retrieval-Augmented Generation (RAG) to provide accurate and explainable colon cancer classification from histopathology images. The proposed system classifies uploaded histopathology tissue images as benign or malignant while generating attention map visualizations that highlight the image regions contributing most to the model's prediction. To strengthen diagnostic transparency, the framework retrieves visually similar histopathology cases from a reference dataset and uses them as supporting evidence for its reasoning process. By combining transformer-based image understanding, visual explainability, and retrieval-based evidence, RAG-ViT enables clinicians to better interpret AI predictions and compare them with relevant historical examples. The proposed approach enhances transparency, interpretability, and confidence in AI-assisted pathology, making it a promising decision-support tool for reliable and trustworthy colon cancer diagnosis.
See Video: https://drive.google.com/file/d/1Vt9sp1Y0IFTudJoG9l-x25G_L9oS4WYr/view?usp=drive_link
Local RAG-Enhanced Small Language Model for Heart Disease Prediction with In-Context Clinical Reasoning on Structured Tabular Data
Researchers have extensively employed traditional machine learning (ML) techniques for the prediction and classification of CVDs. Although existing machine learning models for cardiovascular disease prediction achieve satisfactory accuracy, they often function as black boxes, lacking transparency in their decision-making process. This opacity limits their ability to provide meaningful explanations for predictions, thereby restricting their practical utility and trustworthiness for users and clinicians. To address this limitation, this study proposes a novel explainable model that predicts cardiovascular disease while delivering clear reasoning behind each prediction. This model will remove the attributes which does not make contribution for prediction and it is local and Web-Deployed small language model (SLM) using a structured tabular clinical Cleveland dataset. The system integrates Retrieval-Augmented Generation with a SLM to improve predictive performance with contextual reasoning. It predicts heart disease outcomes while providing transparent explanations to enhance user trust. The model achieves high accuracy on evaluated datasets. To further improve interpretability, SHAP and LIME are applied to visualize feature importance and attribute contributions.
Page updated
Google Sites
Report abuse