Deep Learning -Based Algorithm for MRI Lumbar Vertebrae Instance Segmentation

Magnetic resonance (MR) morphometric analysis of the spine plays crucial role in clinical practice for diagnosis various anomalies, including osteoporosis, age-related degenerative changes, etc. However, the current approach to MR morphometry relies on manual measurements by radiologists, which is a...

Ausführliche Beschreibung

Gespeichert in:
Bibliographische Detailangaben
Veröffentlicht in:2023 IEEE Ural-Siberian Conference on Computational Technologies in Cognitive Science, Genomics and Biomedicine (CSGB) S. 150 - 153
Hauptverfasser: AI-Haidri, Walid, Brui, Ekaterina, Usmanova, Indira, Salimov, Farkhad, Akhatov, Ainur, Al-Habeeb, Mohammed, Il'yasov, Kamil A.
Format: Tagungsbericht
Sprache:Englisch
Veröffentlicht: IEEE 28.09.2023
Schlagworte:
CNN
Current measurement
Deep learning
DL-based Algorithm
Image Segmentation
Magnetic resonance
Magnetic resonance imaging
MRI
Osteoporosis
Spine
Training
vertebra deformity
Online-Zugang:Volltext
Tags: Tag hinzufügen
Keine Tags, Fügen Sie den ersten Tag hinzu!
Beschreibung
Zusammenfassung:Magnetic resonance (MR) morphometric analysis of the spine plays crucial role in clinical practice for diagnosis various anomalies, including osteoporosis, age-related degenerative changes, etc. However, the current approach to MR morphometry relies on manual measurements by radiologists, which is a time-consuming process. Therefore, automating morphometric analysis is important to speed up the measurements and reduce the workload on radiologists. In order to achieve this, automated segmentation of the vertebrae can be implemented. This work proposes a deep learning (DL)-based algorithm for segmenting lumbar vertebrae in magnetic resonance (MR) images. The algorithm utilizes the Mask-RCNN deep convolutional neural network, which is a cutting-edge tool for multiple objects (instance) segmentation. The network was trained and evaluated on a dataset of MR images, consisting of 200 subjects (100 control and 100 patients with at least one deformed vertebrae). These images were carefully labelled. The trained Mask-RCNN model demonstrated exceptional segmentation performance on the test dataset. The median Dice similarity coefficient, a widely used metric for evaluating segmentation accuracy, was 0.95 for patients and 0.96 for controls. The results can be utilized to automate the assessment of lumbar vertebrae deformities, including the identification of deformity type (wedge-shaped or biconcave, etc.), as well as the determination of deformity grade.
DOI:10.1109/CSGB60362.2023.10329828