A Deep Neural Network-based Framework for Cochlear Implant Hearing Compensation

A Deep Neural Network-based Framework for Cochlear Implant Hearing Compensation @WAVES, Information Technology, Ghent University Authors: Julie Van Heghe, prof. dr. Sarah Verhulst, prof. dr. ir. Guy Torfs Cochlear implants have restored hearing to over a million people worldwide, offering benefits for individuals with severe to profound hearing loss. However, current CI systems still face significant limitations due to the simplified and inflexible processing strategies used in modern implants. This research explores the integration of deep learning techniques into the cochlear implant modelling pipeline, with the goal of bringing the cochlear implant output closer to the natural auditory response. [email protected] •Electric pulses stimulate the auditory nerve by an implanted electrode array •Executed in the sound processor of cochlear implants Working of Cochlear Implants Sound waves Electric pulses Figure 1: Internal Components Of A Cochlear Implant | Advanced Bionics [1] By comparing the normal hearing pathway with the cochlear implant-stimulated system, a deep neural network (DNN-CI) is trained to adapt the cochlear implant input in a way that minimizes the difference between the two outputs, effectively mimicking natural hearing. Normal Hearing System Cochlear Implant system Loss function Cochlea Module Inner Hair Cell Module Synapse Module Spiral Ganglion Module Stimulation Module*DNN-CI Current Spreading Module Spiral Ganglion Module Sound waves Cochlear vibrations IHC receptor potentials Postsynaptic currents SGN receptor potentials Sound waves Sound waves Electric pulses (of electrodes) Electric pulses (of neurons) SGN receptor potentials The DNN-CI module has a convolutional neural network-based architecture called ‘dCoNNear’. The current DNN-CI block is unable to sufficiently adjust the HI input such that the resulting SGN output mimics that of the NH pathway: *DNN-CI Results •Improvement of the differentiable models, more specifically the stimulation model, to accurately mimic the behavior of sparse pulses. •Improvement of the computational models, more specifically the spiral ganglion model, to accurately represent the response to fine temporal implant stimulation. Analytical models Converted into deep neural networks to enable backpropagation in the closed loop. Example for the synapse model: Each model is trained to replicate the analytical outputs. Differentiable Modules Work in Progress