SPINESLICER: EXTERNALLY VALIDATED DEEP LEARNING MODELS FOR LUMBAR SPINE SEGMENTATION, GRADING AND DISC HEIGHT INDEX CALCULATIONS

Narasimharao Kowlagi, Antti Kemppainen, Terence McSweeney, Simo Saarakkala, Jerome Noailly, Frances MK Williams, Jason Pui Yin Cheung, Jaro Karppinen, Huy Hoang Nguyen, Aleksei Tiulpin

Abstract

This study introduces deep learning models for automated lumbar spine segmentation and grading with disc height index calculation via a 3D Slicer extension. Leveraging diverse datasets, we enhance model generalizability and accuracy, facilitating efficient exploration of complex phenotypes related to low back pain and their biopyschosocial factors.

Results

Predicitons
Comparison of segmentation performance between the base and adapted models across six external lumbar spine MRI datasets. The adapted model demonstrates improved generalization providing more precise detection of vertebral bodies and intervertebral discs.
Pipeline
Pipeline
Model comparison between our model and the TotalSegmentator model for vertebral body and disc segmentation. Our model demonstrates superior performance across all datasets, with comparable performance on the Twins UK and SPIDER datasets, highlighting its robustness and generalizability.
Comparison of segmentation performance (Dice similarity coefficient; DSC) between the base model, the augmented model, and the adapted model on six external datasets. The base model was trained on the NFBC1966 dataset. Inter-rater agreement (DSC) is also reported for five of the datasets.

Demo

Windows Installation