Application of Deep Learning-Based Medical Image Segmentation via Orbital Computed Tomography

An object segmentation protocol for orbital computed tomography (CT) images is introduced. The methods of labeling the ground truth of orbital structures by using super-resolution, extracting the volume of interest from CT images, and modeling multi-label segmentation using 2D sequential U-Net for orbital CT images are explained for supervised learning.

Transcript

We introduced the object segmentation protocol for orbital computed tomography images in this study. It is the first trial in the world. Using this program, you can do masking on any anatomy part easily where you want to.

It visualizes and saves your time and effort. We hope this study can be a cornerstone for the diagnosis of orbital diseases, which are difficult for biopsy. Begin by running the masking software program.

To load the orbital CT, click the open file icon and select the target CT file. To mask the eyeball, optic nerve, and extraocular muscles using superpixels, run the SmartPencil by clicking on the SmartPencil wizard in MediLabel. Next, control the resolution of the super pixel map, if necessary, and click on the cluster of superpixels of the eyeball, extraocular muscles, and optic nerve on the superpixel map, where pixels of similar image intensity values are clustered.

To refine the masks, click on the SmartFill wizard after masking some of the superpixels on the slices, then click on the AutoCorrection icon, and ensure that the corrected mask labels are computed. Once the refinement of the masking is complete, save the masked images. Run the Python script for pre-processing, and check the scans and masks which are cropped and saved in the VOIs folder.

Run the sequence builder Python script to transform the VOIs into a set of three sequential CT slices to use as the input for sequence U-Net. Check the saved, transformed CT scans and masks in the scan folder and the mask folder and the pre-processed folders respectively. To build the orbital segmentation model, run the Python script main.

py and give the fold numbers. Set the epoch, which is the number of training iterations, and set the batch size, which is the number of training samples in a single training session. The script main.

py can run without the parsers, in which case, it runs with default values. Perform the testing of the model after training, and calculate the evaluation metrics such as dice score and volume similarity. Finally, check the results saved as image files.

The eyeball segmentation using sequence U-Net for orbital structure segmentation achieved a visual similarity or VS score of 0.83 and a high dice score of 0.86 because it had a large portion of the VOIs and little heterogeneity between CT scans. A low dice score score of 0.54 was achieved for the segmentation of the extraocular muscles, and 0.34 for optic nerve, because they infrequently appeared in the CT volume and were found in a relatively small number of CT slices. However, the visual similarity scores of the extraocular muscles and optic nerve were higher than their dice scores, which indicates that the specificity of segmentation was low.

Overall, the segmentation of all the orbital substructures achieved a dice score of 0.79 and a visual similarity score of 0.82. Depending on the application, the size of VOI and the level of window clipping can vary. You can modify the sequence builder code for other great purposes.

Also, the hyperparameters for model training can be modified. The model was trained with 46 VOIs, which is not a large number for model training. To overcome the low performance due to the small number of training data sets, transfer learning and domain notation could be applied.

Read Article

Journal (Engineering)

Journal (Engineering)

Application of Deep Learning-Based Medical Image Segmentation via Orbital Computed Tomography

Chapters

Transcript

Related Videos

Get cutting-edge science videos from JoVE sent straight to your inbox every month.

Your free access has ended.