Skip to main content
AAO-HNSF 2023 Annual Meeting & OTO Experience Main banner
Icon Legend
This session is not in your schedule.
This session is in your schedule. Click again to remove it.
Back
Rate ePoster

Otology/Neurotology

(057) Automated Segmentation of Middle Ear Anatomy for Endoscopic Ear Surgery

Monday, October 2, 2023
2:45 PM – 3:45 PM East Coast USA Time
Has Audio

    Presenting Author(s)

  • Obinna Nwosu, MD

    Resident Physician
    Massachusetts Eye and Ear Infirmary/Harvard Medical School
    Boston, Massachusetts, United States

    • Co-Author(s)

  • KS

    Krish Suresh, MD

    Resident Physician
    Massachusetts Eye and Ear Infirmary/Harvard Medical School
    Boston, Massachusetts, United States

    • Senior Author(s)

  • DL

    Daniel J. Lee, MD

    Massachusetts Eye and Ear Infirmary/Harvard Medical School, USA

  • Matthew G. Crowson, MD

    Assistant Professor
    Massachusetts Eye and Ear Infirmary/Harvard Medical School
    Boston, Massachusetts, United States

Disclosure(s):
Obinna Nwosu, MD: No relevant relationships to disclose.
Krish Suresh, MD: No relevant relationships to disclose.
Matthew G. Crowson, MD: No relevant relationships to disclose.


Introduction: Accurate identification of relevant middle ear anatomy is prerequisite to safe endoscopic ear surgery (EES) and to the development of augmented-reality/robotic-assisted surgical platforms. This study describes a model for automated segmentation of middle ear anatomy during EES.

Methods: We developed a segmentation model that identifies middle ear anatomy using images acquired from videos of endoscopic stapes surgery publicly available online. The model was trained to identify the chorda tympani (CT), incus, malleus, promontory, round window niche (RWN), stapedial tendon (ST), and tympanomeatal flap (TF). We created a ground-truth dataset by manually segmenting relevant anatomic structures from 105 extracted still images. Predicted segmentations were compared against ground-truth segmentations using average precision (AP) and Intersection over Union (IoU).

Results: The model’s overall AP was 55% at an IoU of 50. In other words, for all relevant structures combined, the predicted labels correctly overlapped with the ground-truth labels by at least 50%, 55% of the time. Structure-specific AP were as follows: incus (61%), TF (48%), CT (27%), promontory (26%), ST (20%), RWN (16%), and malleus (8%). Certain structures, such as incus, may be more successfully segmented because they are more easily distinguished from surrounding anatomy. This is contrast to other structures, such as the promontory, whose identification relies on its relation to surrounding structures. Each anatomic structure was not readily visible, and thus not manually segmented, in each image. Structures which were underrepresented in the image dataset, such as the ST, CT, and malleus, tended to be less precisely segmented.

Conclusions: We developed a novel segmentation model with use of a small, heterogenous dataset obtained from online EES videos. To improve model performance, we plan to retrain the model with a larger dataset of higher resolution images from our own surgical videos. Successful automated segmentation of relevant anatomy can help to develop platforms for EES training and augmented-reality/robotic-assisted surgery.