SVseg: Stacked Sparse Autoencoder-Based Patch Classification Modeling for Vertebrae Segmentation

Precise vertebrae segmentation is essential for the image-related analysis of spine pathologies such as vertebral compression fractures and other abnormalities, well clinical diagnostic treatment surgical planning. An automatic objective system vertebra required, but its development likely to run into difficulties low accuracy requirement prior knowledge or human intervention. Recently, methods have focused on deep learning-based techniques. To mitigate challenges involved, we propose learning primitives stacked Sparse autoencoder-based patch classification modeling Vertebrae (SVseg) from Computed Tomography (CT) images. After data preprocessing, extract overlapping patches CT images input train model. The sparse autoencoder learns high-level features unlabeled image in an unsupervised way. Furthermore, employ supervised refine feature representation improve discriminability learned features. These are fed a logistic regression classifier fine-tune A sigmoid added network discriminate non-vertebrae by selecting class with highest probabilities. We validated our proposed SVseg model publicly available MICCAI Computational Spine Imaging (CSI) dataset. configuration optimization, achieved impressive performance, 87.39% Dice Similarity Coefficient (DSC), 77.60% Jaccard (JSC), 91.53% precision (PRE), 90.88% sensitivity (SEN). experimental results demonstrated method’s efficiency significant potential diagnosing treating spinal diseases.