Developed by researchers at the University of Göttingen, the tool examines the mouse cochlea in one piece, without cutting it into sections. In animal experiments it shows, across the whole cochlea, which cells gene and optogenetic therapies reach and whether they cause cell loss.
In this article we look at the study titled "CochleaNet: deep learning-based image analysis for cochlear connectomics and gene therapy", accepted for publication in Molecular Therapy in October 2026 and written by Lennart Roos, Aleyna Miraç Diniz and colleagues. The work was carried out at the University of Göttingen by the teams of Tobias Moser and Constantin Pape.
The cochlea is a spiral organ that sits inside the temporal bone; in the mouse it is only a few millimetres across. Until now its cells have mostly been counted in thin sections, or in small pieces of the organ of Corti dissected out of the cochlea. Part of the tissue can be damaged in the process, and counting is usually done by hand in selected regions, with the whole organ then estimated from those regions. Yet in order to judge a gene therapy you need to see, across the whole cochlea, how many of the target cells the therapy reaches, whether that proportion differs between the basal and apical regions, and whether any cells have been lost.
The cochlea is first softened in a solution that removes calcium, after which the neurons, the inner hair cells and the ribbon synapses are labelled with separate fluorescent markers. A ribbon synapse is the contact at which an inner hair cell passes sound information to a fibre of the auditory nerve. Finally the tissue is made almost as clear as glass by chemical means; this preparation takes six to eight weeks. Imaging is done with the light-sheet microscope that a team at the same university described in Nature Biotechnology in 2025. The microscope scans the specimen layer by layer with a thin sheet of light coming in from the side, and its resolution is the same in every direction, so cells lying very close together can also be separated along the depth axis. One mouse cochlea is scanned in about an hour.
CochleaNet is made up of three separate deep learning models: one finds the spiral ganglion neurons that form the auditory nerve, one the inner hair cells and one the ribbon synapses. Because each synapse is matched to the nearest inner hair cell, the number of synapses on each cell can be worked out one by one. When the results were compared with expert annotations, the model's accuracy in counting neurons came close to the agreement between the experts themselves, while it stayed somewhat behind for the inner hair cells and the synapses. General purpose AI models often used for cell segmentation were reported to have given insufficient results on the same data.
Average of four mouse cochleae (Roos et al., 2026).
These numbers agree with earlier work, although most of them sit near the upper end of the reported range. The authors put this down to there being no cell loss, since the organ is not taken apart, and to the counts resting on the whole cochlea rather than on an estimate.
Because the whole organ is counted, every finding can be placed on a frequency map according to where it sits along the cochlea. The team used the Greenwood function for this.
Every region of the cochlea responds best to a particular frequency: the base to high frequencies, the apex to low ones. The Greenwood function is a formula that converts the position of a point along the cochlea into the frequency that point responds to best; its coefficients differ between species, so separate values are used for humans, mice and gerbils.
The number of synapses per inner hair cell stayed around 11 to 15 across a wide region corresponding to roughly 2 to 64 kHz in the mouse cochlea, and fell only at the two ends. The team also sorted the neurons by subtype. Type I neurons, which carry information from the inner hair cells to the brain, make up the great majority of the auditory nerve and divide further into the Ia, Ib and Ic subgroups; these subgroups are thought to correspond to fibre groups that differ in their sensitivity to sound. Type II neurons, which connect to the outer hair cells, account for only about 3.5 to 5%. The proportions of the subgroups did not change markedly from base to apex.
In gene therapy an intact copy of a missing or faulty gene is carried into the cell by a virus that has been rendered harmless, usually an adeno-associated virus (AAV). In optogenetic therapy the gene carried into the cell makes it produce a light-sensitive protein; optical cochlear implants, which work with light rather than electricity, rest on this principle.
The team delivered the gene for ChReef, the protein used in optical cochlear implant research, to newborn mice through the left ear only. In the left ear about 90% of the neurons were producing the protein, and in the right ear, which was not injected, the figure was 84%. The authors attribute this to the virus reaching the other ear through the cerebrospinal fluid. On the other hand the neuron count fell from 11,176 in untreated mice to 9,954 in the ears that were not injected and to 7,523 in the ears that were. According to the authors this drop may point to a partial loss of neurons caused by the injection, by the virus or by the protein produced. In a study that looks at only part of the cochlea, a problem of this kind could easily go unnoticed.
The team also tried the method in gerbils with another light-sensitive protein and in mice with OTOF gene therapy. No marked neuron loss was seen in the gerbils. The OTOF experiment used only two mice, so it should be read less as a result than as an example of the kind of question the method can answer.
CochleaNet is not a clinical tool, nor is it a treatment in itself. It may, however, speed up laboratory work in two areas that concern audiology closely. The first is gene and optogenetic therapy: OTOF gene therapy is giving promising results in clinical trials in children, and the first clinical trial of an optical cochlear implant is expected to begin within the next few years. Before a treatment is tried in people, animal experiments have to show which cells it reaches and whether it causes cell loss.
The second is cochlear synaptopathy, that is, hidden hearing loss. Animal studies show that noise and ageing can make the synapses of the inner hair cells disappear without any clear change on the audiogram (Kujawa and Liberman, 2009; Sergeyenko et al., 2013). CochleaNet could make it easier to follow this loss across the whole cochlea, region by region. Because the code and the sample data have been shared openly, other laboratories with a light-sheet microscope will be able to run the same analysis on their own specimens.
Note: according to the declaration of interests in the article, Tobias Moser is among the co-founders of OptoGenTech, a company developing optogenetic hearing therapy.
Roos, L., Diniz, A. M., Koert, E., Schilling, M., Uhl, M., Thirumalai, A., Aakhte, M., Kusch, K., Huisken, J., Moser, T., & Pape, C. (2026). CochleaNet: deep learning-based image analysis for cochlear connectomics and gene therapy. Molecular Therapy. https://doi.org/10.1016/j.ymthe.2026.10.025
Aakhte, M., Müller, G. F., Roos, L., Li, J., Göpel, T., Weiss, K. R., Diniz, A. M., Wenzel, J., Schwaninger, M., Moser, T., & Huisken, J. (2025). Isotropic, aberration-corrected light sheet microscopy for rapid high-resolution imaging of cleared tissue. Nature Biotechnology. https://doi.org/10.1038/s41587-025-02882-8
Alekseev, A., Hunniford, V., Zerche, M., et al. (2025). Efficient and sustained optogenetic control of sensory and cardiac systems. Nature Biomedical Engineering, 10, 277-292.
Keppeler, D., Kampshoff, C. A., Thirumalai, A., et al. (2021). Multiscale photonic imaging of the native and implanted cochlea. Proceedings of the National Academy of Sciences, 118, e2014472118.
Lv, J., Wang, H., Cheng, X., et al. (2024). AAV1-hOTOF gene therapy for autosomal recessive deafness 9: a single-arm trial. The Lancet, 403, 2317-2325.
Qi, J., Zhang, L., Lu, L., et al. (2025). AAV gene therapy for autosomal recessive deafness 9: a single-arm trial. Nature Medicine, 31, 2917-2926.
Kujawa, S. G., & Liberman, M. C. (2009). Adding insult to injury: cochlear nerve degeneration after "temporary" noise-induced hearing loss. Journal of Neuroscience, 29(45), 14077-14085.
Greenwood, D. D. (1990). A cochlear frequency-position function for several species: 29 years later. The Journal of the Acoustical Society of America, 87(6), 2592-2605.
Sergeyenko, Y., Lall, K., Liberman, M. C., & Kujawa, S. G. (2013). Age-related cochlear synaptopathy: an early-onset contributor to auditory functional decline. Journal of Neuroscience, 33(34), 13686-13694.
Auditory Scene