Year: 2025
Type: article
Abstract: Medical imaging depends heavily on spine segmentation for diagnosing spinal disorders as well as developing treatment plans. The proposed research describes a deep learning framework for spine segmentation where two state-of-the-art models known as DeepLabv3+ and U-Net are evaluated for comparison. Our team applied the developed framework using 1089 high-resolution CT images which had been annotated with segment outlines for cervical, thoracic, lumbar and sacral areas. Person working with the DeepLabv3+ model produced segmentations with 0.925 Intersection over Union (IoU) and 0.920 Dice score along with 0.920 F1 score and an accuracy rate of 98.5%. The U-Net model achieved better performance when applied with a ResNet34 backbone by reaching an IoU of 0.953 alongside Dice score 0.951 F1 score 0.951 and accuracy reaching 98.6%. U-Net outperforms existing methods of spine segmentation by achieving superior results in area detection along with boundary definition accuracy. The research provides a thorough assessment of network designs and practical utility tests to support the development of automated spine segmentation applications that can work effectively in clinical settings.
Language: English
Topic: Medical Imaging and Analysis
Subfield: Biomedical Engineering
Field: Engineering
Domain: Physical Sciences
Open Access status: gold
Estimated APC paid: $500
OpenAlex ID: W4415091653