Purpose or Case Report: There is growing research in diffusion tensor imaging (DTI) for providing metrics and images of physeal structure and function, particularly in evaluating children with possible growth disorders. There is a need for a faster automated process to segment the diffusion data, however. We utilized a deep learning algorithm to automatically generate a region of interest (ROI) for use in fully automatic diffusion tensor tractography of cartilage columns in the distal femoral physis.
Methods & Materials: Following IRB approval, the authors retrospectively analyzed 80 knee DTI studies from 40 children including 20 neuroblastoma survivors with growth failure and 20 matched controls.
Manual binary segmentation of the femoral physis on B0 images was performed on all 80 sequences. Saved volumes were automatically cropped to a 40x24x40 voxel 3-dimensional ROI including the distal femoral physis, femoral condyles, and proximal tibia. A novel, fully connected CNN based on a 3D modification of the U-Net architecture was trained on the output volumes (n=70) to generate segmentation masks of only the distal femoral physis. A holdout set of B0 sequences (n=10) was segregated prior to training to serve as the testing set.
Training data augmentation mainly included random rigid affine rotation of the input volumes about the 3 axes and simulated Gaussian noise generation. The segmentation network was trained for 600 epochs. Software code was written in Python v3.6 using the TensorFlow v1.13 module on a Linux workstation with two NVidia Titan X GPUs.
Results: The overall mean Dice correlation coefficient was 0.680 with a Matthews Correlation Coefficient of 0.672. Manual extraction and segmentation took 10 minutes per volume and had an interobserver correlation of 0.82, whereas DL segmentation took <1 second per volume and was deterministic, always producing the same result for a given input.
Conclusions: Deep learning (DL) techniques can be applied to B0 sequences in DTI examinations in order to generate a deterministic ROI of the distal femoral physis. DL automates physeal mapping, increasing speed, and decreasing variability of tractography.