Computed tomographic radiography (CT) can be potentially used for knee osteoarthritis (OA) diagnosis and follow up. It is possible to quantify knee joint space three-dimensionally. The main purpose of this study was to investigate an accurate, precise, and memory-saving convolutional neural network to accomplish the segmentation of lateral joint space and medial joint space (LJS and MJS) of knee joints. A novel deep learning structure, based on two state-of-the-art networks — RegNet and DeepLabV3P, was constructed for automatic prediction. And a new attention mechanism called "Feature-Location" module was proposed to be added into the models for comparison. The predictions were post-processed by area opening and alpha shape to get the complete segmentations. From the joint space segmentations, morphological measurements clinically pertinent for knee OA diagnosis and follow up such as the volume, mean thickness, and standard deviation were measured. Finally, compared with the standard U-Net and DeepLabV3P, our new model called "DeepRegY" reduced the memory footprint by more than double. The model also produced the highest Dice coefficients, 0.8378 for LJS and 0.8655 for MJS in the validation set. The morphological parameters showed strong correlations, 0,9356 for volume, 0.99 for mean thickness, and 0.9876 for standard deviation of thickness in the testing set. Bland-Altman analysis showed a mean thickness bias about -0.2 mm which is very low compared to the mean thickness results.