Authors:
Aseem Jain (Iowa City, IA | US)
Nicholas George-Jones (Iowa City, IA | US)
Rachel Scheperle (Iowa City, IA | US)
Joshua Pinzour (Iowa City, IA | US)
Marlan Hansen (Iowa City, IA | US)
Alexander Claussen (Pomerantz Family Pavilion, Iowa City, IA | US)
Goals
Intracochlear cochlear implant (CI) electrode array (EA) position has been known to impact post-operative hearing outcomes; serving as measure for CI trauma assessment. Current tools for CI trauma assessment have limitations including significant user input, restricted to proprietary EAs, and lack of distance metrics between EA and intracochlear structures. Our aim was to create an automated, manufacturer agnostic tool (ICAT) that segments the EA and cochlea within post-operative CT scans (poCT), and overlays labels for the scala-tympani (ST), scala vestibuli/media (SVM), lateral wall (LW), modiolus, and basilar membrane (BM) to compute EA position relative to these intracochlear structures.
Material and Methods:
The tool uses two pipelines: 1) Cochlea and EA segmentations derived from poCT; 2) Intracochlear segmentations (listed above) derived from pre-operative CT. Pipeline 1 was created by training a neural network (nnUNet) on 50 labeled poCTs from CI patients. Pipeline 2 was developed using five co-registered 3D-xray microscopy and CT pairs of the cochlea from cadaveric temporal bones, augmented synthetically to train a separate nnUNet. The result of pipeline 2 is registered to pipeline 1 for EA position analysis. Cochlear duct length (CDL) is estimated by computing a centerline through ST generated from pipeline 2. For both pipelines, using five-fold cross validation, 20% of the training data were randomly isolated to create a testing dataset. Both pipelines were integrated within the open-source platform, Slicer3D.
Results:
Pipeline 1 achieved Dice similarity coefficients (DSC) of 0.94 for cochlea segmentation and 0.87 for EA segmentation. Pipeline 2 achieved DSCs of 0.96, 0.95, 0.83, 0.86, and 0.91 for ST, SVM, BM, modiolus, and LW respectively. On a subset of LW CIs (N=6), ICAT found that, on average, 89.7% EA was in the ST, 6.3% in LW, and 3.4% SVM. Within this subset, mean EA distances to BM, LW, and modiolus were 0.50[-0.42-1.26] mm, 0.61[0.03-1.63] mm, and 1.26 [0.4-2.91]mm, respectively. Additionally, ICAT captures distance metrics as a function of percent of CDL (0% = base, 100%=apex); the average minimum distance to BM, LW, and modiolus occurred at 62%, 45%, 36% of the CDL.
Conclusion:
We created an accurate, fully automated CI trauma assessment tool that is EA-agnostic. This platform facilitates reproducible, large-scale analysis to guide surgical and programming strategies aimed at improving CI outcomes.