The Dresden in vivo OCT dataset for automatic middle ear segmentation

Area of Science:

• Biomedical Imaging
• Medical Informatics
• Otolaryngology

Background:

• Endoscopic optical coherence tomography (OCT) enables non-invasive in vivo assessment of the middle ear.
• Interpreting middle ear OCT images is complex due to shadowing, hindering clinical application.
• Deep neural networks show potential for OCT image analysis but require large annotated datasets.

Purpose of the Study:

• To introduce the Dresden in vivo OCT Dataset of the Middle Ear (DIOME).
• To provide a benchmark dataset for developing and evaluating AI algorithms for middle ear OCT analysis.
• To facilitate automated diagnostics and morphological/functional assessment of the middle ear.

Main Methods:

• Collected 43 OCT volumes from 29 subjects with healthy and pathological middle ears.
• Generated semantic segmentations for five key anatomical structures: tympanic membrane, malleus, incus, stapes, and promontory.
• Annotated salient features with sparse landmarks.

Main Results:

• The DIOME dataset comprises 43 annotated OCT volumes.
• Provides detailed semantic segmentations and landmark annotations for crucial middle ear structures.
• Enables robust training and evaluation of deep learning models for middle ear OCT analysis.

Conclusions:

• DIOME addresses the scarcity of annotated middle ear OCT datasets.
• Facilitates advancements in AI-driven diagnostics for middle ear pathologies.
• Supports research in automated image analysis for otolaryngology.