Deep learning automatische Klassifikation von Innenohrdysplasien basiert auf lateral wall Anatomie – erste Ergebnisse

Aim:

The anatomical heterogenicity of inner ear dysplasia make their identification and classification notoriously challenging. As the surgical complications to be expected and the cochlear implantation results tend to correlate with the underling anatomical anomaly, their proper identification remains paramount. This study aimed to develop an artificial intelligence (AI) model for the automated recognition of cochlear dysplasia using the lateral wall anatomy.

Methods:

A segmentation database comprising 141 normal cochleae and 117 dysplastic cochleae (12 Cochlea Hypoplasia Type 2, 17 Cochlea Hypoplasia Type III, 40 Incomplete Partition Type I, 40 Incomplete Partition Type II, 8 Incomplete partition Type III) was created. The AI algorithm was trained to differentiate among normal and dysplastic cochleae and to classify these using the lateral wall anatomy.

Results:

The multivariate analysis identified the radius of the second cochlear turn as an important parameter to characterize different cochlea dysplasia. The AI model demonstrated good performance in distinguishing cochlear dysplasia from normal cochleae and recognising the dysplasia subtype. Precision ranged from 0.92 to 1.00, recall from 0.79 to 1.00, and the F1-score from 0.76 to 1.00.

Conclusions:

These results underscore the efficacy of integrating AI into diagnostic workflows, enabling accurate and efficient recognition of cochlear dysplasia. Automated recognition of cochlear dysplasia based on lateral wall anatomy could streamline diagnostics, reduce human diagnostic error, and improve accessibility to expert-level evaluations. Furthermore, this advancement paves the way for personalized therapeutic strategies tailored to the specific dysplasia subtype, enhancing rehabilitation outcomes.

Nein