37 - Semi-automated 3D Tissue Segmentation for Whale Larynges from MRI Studies

Poster Board Number: 37
There are separate poster presentation times for odd and even posters.
Odd poster #s – first hour
Even poster #s – second hour

Co-authors:

Joy Reidenberg, PhD - PROFESSOR, MEDICAL EDUCATION, Icahn School of Medicine at Mount Sinai; Jonathan Wisco, PhD - ASSOCIATE PROFESSOR, ANATOMY & NEUROBIOLOGY, Chobanian and Avedisian School of Medicine

Abstract Body : Introduction:

To elucidate the roles that a whale larynx may play in sound generation, we aim to build and characterize an atlas of mysticete (baleen whale) anatomy. This characterization may be achieved through 3D segmentation of MRI scans of whale laryngeal tissue. While manual segmentation may be the standard for tissue quantification, the task of hand-segmenting thousands of slices over many whale larynges proves to be laborious and expensive to complete. We hypothesize that a semi-automated method for 3D segmentation of MRI images can expedite and reduce the cost of quantifying whole whale larynges while maintaining the integrity of whale tissue segmentation.



Methods:

Whale larynges were collected post mortem from 2 adult male specimens representing 2 species of baleen whales: sei whale, Balaenoptera borealis, and fin whale, B. physalus. Specimens were scanned using a Siemens 3T Skyra MRI scanner with a slice thickness of 1.1 mm, pixel dimensions of 432 x 512, and pixel spacings of 0.86 mm x 0.86 mm. Axial T1-weighted MRI images were used for segmentation. All image processing algorithms were written in Python using the VTK package, and results were visualized using 3D Slicer. The semi-automated segmentation methodology included an image pre-processing step along with manual selection of a threshold value to separate whale tissue from background noise. Due to the size of the whale larynges, imaging artifacts in the peripherals of the MRI stacks were noted and removed algorithmically by selecting for the largest contiguous volume.



Results:

For our preliminary results, we measured the time required to complete the semi-automated 3D segmentation methodology. Aside from the several seconds of user input time to manually select a threshold, our semi-automated method averaged a total time of 0.273 ± 0.017 seconds to preprocess, threshold, and clean roughly 100 single axial slices of whale larynx MRI. Timing results were measured using a Ryzen 5600x CPU.



Conclusion:

Moving forward, we aim to continue refining our semi-automated segmentation methodology to include the differentiation of specific tissue types and the creation of a fully-automated approach. A fully-automated approach may include automatic threshold selection or use a deep learning convolutional neural network. Additionally, our collection of whale larynges includes T2-weighted MRI images and CT images, which may be more amenable to segmentation of certain tissue types, e.g., CT for bone. We intend to utilize our 3D segmentation methodology to analyze the whale laryngeal tissues in our collection with the ultimate goal of building an atlas of mysticete anatomy. Objective quantification of these whale laryngeal tissues will aid in untangling the mechanisms of whale vocalization.