Assessment of Optimal Deep Learning Configuration for Vertebrae Segmentation from CT images

摘要

Automated segmentation of vertebral bone from diagnostic Computed Tomography (CT) images has become animportant part of clinical workow today. There is an increasing need for computer aided diagnosis applicationsof various spine disorders including scoliosis, fracture detection and even automated reporting. While modelbasedmethods have been widely used, recent deep Learning methods have shown a great potential in this area.However, choice of optimal conguration of the network to get the best segmentation performance is challenging.In this work, we explore the impact of different training and inference options, including dimensions, activationfunction, batch normalization, kernel size, filters, patch size and patch selection strategy in U-Net architecture. 20publicly available CT Spine datasets from Spineweb repository was used in this study divided into training/testdatasets. Training with dierent DL congurations were repeated with these datasets. We used the best weightscorresponding to each conguration for inference on the independent test dataset. These results on the testdataset with the best weights for each configurations were compared. 3D models performed consistently betterthan 2D approaches. Overlapped patch based inference had a big impact on enhancing performance accuracy.The selection of training patch size was also found to be crucial in improving the model performance. Moreover,the need for an effective balance of positive and negative training patches was found. The best performance in ourstudy was obtained by using overlapped patch inference, training with RELU activation & batch normalizationin a 3D U-Net architecture with training patch size of 128×128×32 that resulted in average values of precision=97%, sensitivity= 96% and F1 (Dice)= 96% for the test dataset.