A Self‐Configuring Deep Learning Network for Segmentation of Temporal Bone Anatomy in Cone‐Beam CT Imaging

Abstract

ObjectivePreoperative planning for otologic or neurotologic procedures often requires manual segmentation of relevant structures, which can be tedious and time‐consuming. Automated methods for segmenting multiple geometrically complex structures can not only streamline preoperative planning but also augment minimally invasive and/or robot‐assisted procedures in this space. This study evaluates a state‐of‐the‐art deep learning pipeline for semantic segmentation of temporal bone anatomy.Study DesignA descriptive study of a segmentation network.SettingAcademic institution.MethodsA total of 15 high‐resolution cone‐beam temporal bone computed tomography (CT) data sets were included in this study. All images were co‐registered, with relevant anatomical structures (eg, ossicles, inner ear, facial nerve, chorda tympani, bony labyrinth) manually segmented. Predicted segmentations from no new U‐Net (nnU‐Net), an open‐source 3‐dimensional semantic segmentation neural network, were compared against ground‐truth segmentations using modified Hausdorff distances (mHD) and Dice scores.ResultsFivefold cross‐validation with nnU‐Net between predicted and ground‐truth labels were as follows: malleus (mHD: 0.044 ± 0.024 mm, dice: 0.914 ± 0.035), incus (mHD: 0.051 ± 0.027 mm, dice: 0.916 ± 0.034), stapes (mHD: 0.147 ± 0.113 mm, dice: 0.560 ± 0.106), bony labyrinth (mHD: 0.038 ± 0.031 mm, dice: 0.952 ± 0.017), and facial nerve (mHD: 0.139 ± 0.072 mm, dice: 0.862 ± 0.039). Comparison against atlas‐based segmentation propagation showed significantly higher Dice scores for all structures (p < .05).ConclusionUsing an open‐source deep learning pipeline, we demonstrate consistently submillimeter accuracy for semantic CT segmentation of temporal bone anatomy compared to hand‐segmented labels. This pipeline has the potential to greatly improve preoperative planning workflows for a variety of otologic and neurotologic procedures and augment existing image guidance and robot‐assisted systems for the temporal bone.

Deep Learning for Automated Image Segmentation of the Middle Ear: A Scoping Review

“…Most recently, Ding et al describe the use of the nnU-Net architecture to perform automated segmentation on 15 CBCTs. 12 nnU-Net, based on a U-net backbone, can automatically self-configure to new data sets. It achieves this by automatically tuning hyperparameters to configure to new data sets.…”

“…Despite the DSC for the malleus of 0.914 ± 0.03 and 0.916 ± 0.034 for the incus, the DSC for the stapes was only 0.560 ± 0.106. 12 Modified Haudorff distance (mHD) was 0.044 ± 0.024 mm for the malleus, 0.051 ± 0.027 mm for the incus and 0.147 ± 0.113 for the stapes. Constrained linear regression between DSC and mHD showed that DSC for small structures like the stapes decreased significantly compared to mHD (p < 0.0001) versus larger or thicker structures.…”

“…DSC was the most commonly reported outcome measure among the included studies. 7,10,12,[14][15][16][17][18][19]…”

“…Only 3 studies broke the ossicular chain down into its components, 12,17,19 whereas all other studies reported a DSC for the ossicles as a whole. For the studies which individually performed stapes segmentation, when compared to Powell et al's atlas-based approach 23 which found an average DSC of 0.53, CNN segmentation in all included papers gave rise to superior performance metrics, with the lowest DSC for the stapes being 0.56 12 to the highest 0.84 using 3D-V-Net. 19 The highest DSC for any middle ear structure was for the incus also using 3D-V-Net at 0.93.…”

Deep Learning Techniques and Imaging in Otorhinolaryngology—A State-of-the-Art Review

“…Highly accurate automatic segmentation of radiological images has been demonstrated in CT scans [107], MRI scans [108] and cone-beam scans [109,110]. After selecting the region of interest, deep learning models can be specifically trained to classify specific diseases, such as chronic otitis media [111], cholesteatoma [112,113], otosclerosis [114,115], mastoiditis [116,117], and Meniere disease [118,119], achieving detection results comparable to subspeciality-trained radiologists.…”