The Dresden in vivo OCT data set for automatic middle ear segmentation

Introduction: Endoscopic optical coherence tomography (OCT) enables a non-invasive high-resolution morphological and functional assessment of the middle ear in vivo. Yet, he interpretation of these OCT images remains challenging and time-consuming due to artifacts and shadowing effects of bony structures. Deep neural networks offer the ability to enhance this process in multiple aspects, including segmentation, classification, and registration. Nevertheless, the scarcity of annotated datasets of OCT middle ear images poses a significant hurdle to the performance of neural networks.
Materials and methods: We utilized a custom endoscopic OCT system with a swept source laser in the range of 1300 nm. The endoscope has a diameter of 3.5 mm. 43 OCT volume scans from both healthy and pathological middle ears of 29 subjects were included in the data set. For each sample, five anatomic structures, including tympanic membrane, malleus, incus, stapes suprastructure, and promontory, were segmented by three different segmentators. Training of the neural network (nnUnet) was done in three iterations, including pathologic samples in the second and third iteration. Technical validation was done by calculation oft he F1 scores and the Hausdorff distance.
Results: The data set is stored at OpARA. The F1 scores were highest fort he Tympanic membrane and lowest for the stapes suprastructure and promontory. The Hausdorff scores were highest for the Malleus.
Conclusion: The Dresden in vivo data set offers a data set of both healthy and pathological OCT volumes oft he middle ear. Thus, it facilitates the training and evaluation of algorithms regarding various analysis tasks with middle ear OCT images, e.g. diagnostics.