In this paper, DNN training and regularization techniques have been explored for lumbar vertebrae segmentation in CT images of patients with degenerative lumbar spine disease. Combinations of loss function, learning rate schedule, normalization, dropo...
In this paper, DNN training and regularization techniques have been explored for lumbar vertebrae segmentation in CT images of patients with degenerative lumbar spine disease. Combinations of loss function, learning rate schedule, normalization, dropout, moving average, weight decay, and squeeze-and-excitation were empirically tested. The results confirmed that the DNN with optimized training and regularization techniques can improve lumbar vertebrae segmentation without complex structural changes.