AD-MAGNet Liver Ultrasound Dataset
Description
This dataset accompanies AD-MAGNet (Agent-Driven Multimodal Graph Attention Network), a deep-learning system for automated liver ultrasound diagnosis. The model classifies liver ultrasound scans into three clinically meaningful categories — Normal, Benign, and Malignant — on the 18,000-image held-out test set. Liver cancer is the sixth most common cancer and third leading cause of cancer-related deaths worldwide. Early, accurate distinction between Normal liver parenchyma, Benign lesions (haemangioma, cysts, focal nodular hyperplasia), and Malignant tumours (hepatocellular carcinoma, cholangiocarcinoma, metastases) is critical for patient outcomes. Traditional radiologist interpretation is time-consuming, subjective, and limited by expertise availability — particularly in low-resource Indian clinical settings. AD-MAGNet addresses this gap through a fully automated, explainable multi-agent AI pipeline. Inspiration & Use Cases Liver tumour detection — train binary or multi-class classifiers directly on the metadata features Biomarker correlation — explore which LFT values best separate the three classes Transfer learning benchmark — use the labelled image paths as ground truth for ultrasound model fine-tuning Multimodal fusion — combine imaging metadata with clinical tabular features Explainability research — uncertainty quantification, Grad-CAM visualisation Indian population studies — demographic-specific liver disease analysis
Files
Steps to reproduce
Liver Ultrasound Dataset — Normal / Benign / Malignant Classification