AutoSpineAI: Lightweight Multimodal CAD Framework for Lumbar Spine MRI Assessments

Saied Salem, Afnan Habib, Mukhlis Raza, Zaid Al-Huda, Omar Al-maqtari, Bilal Ertuğrul, Özal Yıldırım, Yeong Hyeon Gu, Mugahed A. Al-antari

Confirmation: I have read and agree with the IEEE BHI 2025 conference submission's policy on behalf of myself and my co-authors.

Keywords: Lumbar spinal stenosis (LSS), Computer-aided Diagnosis (CAD), Advanced Agentic RAG, Large Language Model (LLM), Structured Medical Report Generation (sMRG)

TL;DR: A Multimodal CAD Framework for Lumbar Spine MRI Assessments and structured Medical Report Generation (sMRG)

Abstract: Automated spine lumbar MRI analysis improves clinical workflow and diagnostic accuracy for lumbar spinal stenosis (LSS). In this paper, we introduce AutoSpineAI, a novel fully automated CAD framework for lumbar spine MRI analysis and structured medical report generation (sMRG) leveraging large language models (LLMs). The system processes 3D MRI DICOM volumes by extracting mid-sagittal slices for vertebrae and intervertebral discs (IVDs) segmentation and localizes corresponding axial slices using 3D cross-projection algorithm. For sagittal and axial slices segmentation, a novel lightweight efficient compact model (ECM) is proposed by integrating multi-attention mechanisms within a compact AI architecture to extract the quantitative spinal structural measurements (SSM): disc degeneration, vertebral anomalies, and other alignment irregularities. These structured measurements and assessments are integrated and merged in prompts for a novel hybrid agentic LLM-driven retrieval system that combines semantic information and knowledge graph-based reasoning to generate detailed level-wise diagnostic report: vertebrae and IVDs. AutoSpineAI achieves Dice scores of 97.58% and 94.01% for sagittal and axial segmentation, respectively, and generates a structured full report by Gemma3 LLM within 30 seconds per patient, achieving 83.51% Bert F1-score, 19.33% Meteor, and 15.31% Rouge1. AutoSpineAI seems to be a scalable and interpretable for clinical and practical solutions for MRI LSS.

Track: 3. Imaging Informatics

Registration Id: K6N34SPB392

Submission Number: 79