*2.2. Image Analysis*

A cooperative multi-agent reinforcement learning framework (C-MARL) as described in [32] was used to automatically detect the cochlear apex, center and round window landmarks for each image. Although only the cochlear center landmark was required for further processing, using a three landmark C-MARL approach ensured better detection of the landmark [33]. CT image-detected cochlear center landmark coordinates, cochlear side and operative status for each CT image were compiled and uploaded to Nautilus (v20220801; Oticon Medical, Vallauris, France)—a web-based cochlear image analysis tool [31].

Nautilus processed the images automatically, generating the cochlear view, intracochlear segmentations and various clinically relevant cochlear parameters. Figure 1 depicts different parameters that Nautilus extracts from each image. Once all the images had been processed, an export bundle was prepared with the following characteristics for analysis: cochlear and ST models, cochlear size, shape, duct lengths and cross-sectional measurements. Nautilus' output confidence scores were also exported and used to filter out any processing failures.

**Figure 1.** Description of clinical metrics computed by Nautilus. A: maximum length between round window and lateral wall; B: maximum perpendicular length to A. h: height of cochlea; h: maximum vertical ST height; r: radius of maximum circle that can fit in ST; ST: scala tympani; SV: scala vestibuli; BM: basilar membrane; OC: organ of corti; LW: lateral wall; MW: modiolar wall.

#### *2.3. Statistical Analysis*

A histogram of the cochlear parameters extracted by Nautilus such as volume, A, B, height, lateral wall (LW) length, the wrapping factor and roller coaster height was generated using 50 bins. Based on the mean and the standard deviation of the parameters, Gaussian curves were plotted on top of the histograms. Correlation analysis was performed via visual inspection of scatter plots and the calculation of the Pearson correlation coefficient between the aforementioned parameters, as well as between parameters A, B and the LW length at various cochlear angles. A regression curve was fitted to the correlation data by the ordinary least squares method. For the correlation and regression analysis, the relevant functions from the SciPy and Scikit-learn python packages were used [34,35].

Analysis of ST height, area and radius was performed up to a cochlear angle of 705◦. The mean, standard deviation, 10th and 90th percentile of the ST angular data were calculated based on the data points falling within ±15◦ of every 30◦ ST angle, e.g., the metadata at 90◦ were based on individual data points between 75◦ and 105◦.

Additionally, an intra-patient analysis was conducted to determine the similarity between contralateral ears. Four hundred fifty-eight patients for whom CT imaging was conducted for both ears were selected for the analysis. The ears were assessed with respect to both imaging and clinical metrics. For imaging analysis, the 3D left–right segmentation meshes were registered together based on landmarks [36,37]. Intra-patient Dice coefficients, Hausdorff distances and average symmetric surface distances were computed [38]. An interpatient analysis was also conducted in which 18 patients were uniformly and randomly selected from the dataset and compared with all other patients (*n* = 440) in the dataset. Global metrics defining cochlear size and shape such as A, B, volume and duct lengths were also evaluated. Statistical *t*-tests with Holm–Sidak correction were performed to analyze the results. A *p*-value of <0.05 was considered significant. A correlation analysis was also performed to determine the relationship between different parameters.

Inter-sex comparison was also carried out based on the cochlear parameters generated by Nautilus. Both size and shape parameters were analyzed to gain insights into whether a distinction could be observed between both sexes. An independent two-sample *t*-test was conducted to determine whether the difference was significant. A *p*-value of <0.05 was considered significant.
