Efficient 3D Deep Learning Model for Medical Image Semantic Segmentation
Nasser Alalwan, Amr Abozeid, AbdAllah A. ElHabshy, Ahmed Ibrahim Alzahrani
Abstract
Medical image segmentation is important for disease diagnosis and support medical decision systems. The study proposes an efficient 3D semantic segmentation deep learning model “3D-DenseUNet-569” for liver and tumor segmentation. The proposed 3D-DenseUNet-569 is a fully 3D semantic segmentation model with a significantly deeper network and lower trainable parameters. The proposed model adopts Depthwise Separable Convolution (DS-Conv) as opposed to traditional convolution. The DS-Conv significantly decreases GPU memory requirements and computational cost and achieves high performance. The proposed 3D-DenseUNet-569 utilizes DensNet connections and UNet links, which preserve low-level features and produce effective results. The results of experimental study on the standard LiTS dataset demonstrate that the 3D-DenseNet-569 model is effective and efficient with respect to related studies.