Raoul Seiler

Semester Work
Supervisors: Prof. Ender Konukoglu

Loss of otoconia: Development of a machine learning algorithm to segment the maculae

Millions people are worldwide affected by, so far unexplained, balance problems. Some come with age, other’s manifest as posttraumatic or residual dizziness. Evidently, the extensive interest and work on sensation of gravity for stable gait and stance, an important and fundamental feature or humans and animals is causing no surprise. Many theories exist how imbalance increases with age, although nobody ever described the loss of otoconia or lack of otoconia as a major reason for imbalance. We tackle this recently made correlation of otoconial loss and imbalance with a first step by developing a convolutional neural network to produce volumetric segmentations of the utricle in the inner ear of humans. The network consisting of a contracting analysis path and a expanding synthesis path with a softmax activation function provides the features of interest. From the input of the network, as plain MRI 3D volumes, to the output of full-resolution segmentations 3D convolutions, 3D max-pooling and 3D up-scaling is used. The output of MRI scans and their segmentations could add significant information to the definiition of otoconial volume in patients and therefore be useful in assisting clinical diagnosis.