Towards Investigating Residual Hearing Loss: Quantification of Fibrosis in a Novel Cochlear OCT Dataset

AI summary

Overview

Research area: Medical image analysis / computer vision applied to cochlear imaging, specifically semantic segmentation of optical coherence tomography (OCT) scans of guinea pig cochleae, with a biomedical motivation in cochlear implant research.

Technical level: Advanced. The paper assumes familiarity with convolutional neural networks, Transformer architectures, encoder-decoder segmentation designs, loss functions, and class-imbalanced segmentation evaluation metrics.

Scope: The paper introduces a novel, manually annotated OCT dataset of chronically implanted guinea pig cochleae with intracochlear fibrosis, benchmarks six contemporary segmentation models on it, and presents a modified 2D UNET (2D-OCT-UNET) as the best-performing approach for quantifying fibrotic burden.

What This Paper Is About

Cochlear implants restore hearing, and hybrid implants that combine electrical stimulation with residual low-frequency acoustic hearing (electroacoustic stimulation, EAS) can offer better speech-in-noise and musical perception. However, up to 50% of EAS patients lose their residual hearing in months to years after implantation, and one suspected contributor is fibrosis — scar-like tissue that forms inside the cochlea in response to the implant. This paper asks whether deep learning can automatically and reliably measure that fibrotic tissue in high-resolution OCT images of implanted rodent cochleae, replacing slow and subjective manual measurement.

Key Contributions

1. A first-of-its-kind annotated dataset. The authors curate a fully annotated Cochlear OCT dataset from five chronically implanted guinea pigs (volumes OCTV1L, OCTV7L, OCTV9L, OCTV10L, OCTV11L), containing 173 annotated slices across three semantic classes plus background. They state this is the first dataset of its kind to capture CI-induced intracochlear fibrosis in a chronically implanted rodent model using OCT.

2. A new task formulation. The paper introduces cochlear semantic segmentation as a task for studying and quantifying cochlear fibrotic burden, defining a fibrosis-percentage equation based on the segmented Fibrosis and ST/Free Space class areas.

3. A dedicated model, 2D-OCT-UNET. The authors design a seven-block encoder-decoder UNET variant with Group Normalization, 37 Conv2D layers with 3×3 kernels, and 1024×1024 input resolution, trained from scratch without transfer learning.

4. A six-model benchmark. They compare their model against VGG16-UNET, UEfficientNet, SegFormer, MST-DeepLabv3+, and the Segment Anything Model (SAM) with adapters, using Accuracy, Precision, Recall, Dice Score Coefficient (DSC), and Jaccard index.

Main Findings

• 2D-OCT-UNET ranked first on four of five metrics. It achieved Accuracy 99.09%, Precision 89.70%, Recall 88.78%, Dice (DSC) 0.8874, and Jaccard 0.8158, with 34,623,099 parameters and a model size of 132.08MB.

• The closest competitor was VGG16-UNET. It reached Accuracy 99.06%, Precision 89.87% (slightly higher than the proposed model's Precision), Recall 87.47%, DSC 0.8788, and Jaccard 0.8061, with 25,862,532 parameters and a 98.66MB model size.

• UEfficientNet placed third by DSC. Accuracy 99.03%, Precision 89.52%, Recall 87.38%, DSC 0.8783, Jaccard 0.8029, with 113,890,355 parameters and a 434.46MB model size.

• SAM with adapters did not lead. Accuracy 98.71%, Precision 84.90%, Recall 85.24%, DSC 0.8402, Jaccard 0.7541, with 94,008,145 parameters and a 358.614MB model size. The paper reports that adding adapters improved SAM somewhat but still left it below 2D-OCT-UNET.

• SegFormer was the weakest by DSC despite being the smallest. Accuracy 98.60%, Precision 85.66%, Recall 79.34%, DSC 0.8080, Jaccard 0.7101, with 3,715,684 parameters and a 14.17MB model size. Its DSC gap to 2D-OCT-UNET was 0.0794, and qualitative figures show pixelated "edge" artefacts in its segmentation maps.

• MST-DeepLabv3+ had the lowest precision and second-lowest DSC. Accuracy 98.57%, Precision 88.99%, Recall 78.07%, DSC 0.8085, Jaccard 0.7204, with 6,407,940 parameters and a 24.44MB model size.

• Fibrosis quantification tracked manual measurement well. Comparing AI-computed and human-computed fibrosis across Z positions, the authors report that AI performance was "excellent" and provided a high-resolution assessment of fibrosis in the scala tympani.

• No significant correlation between absolute error and amount of fibrosis. Absolute error statistics were OCTV1L ρ = −0.1088, p = 0.3464 and OCTV7L ρ = 0.0374, p = 0.8723.

• A small, non-significant negative correlation for signed error. OCTV1L ρ = −0.0857, p = 0.4585 and OCTV7L ρ = −0.1343, p = 0.5616, which the authors interpret as a tendency to underestimate fibrotic burden as it declines in percentage volume of the scala tympani.

• A ground-truth outlier was tolerated. An annotation mistake in the OCTV7L volume, which would have produced a significant p-value of 0.0050 if included in the formal error analysis, was excluded and plotted separately; 2D-OCT-UNET's behaviour on it suggested robustness to certain levels of annotation noise.

• Accuracy was lower for OCTV1L than OCTV7L. The authors attribute this to more noisy labels in that volume and a more imbalanced distribution of semantic class annotations.

Methodology in Plain English

The researchers implanted guinea pigs with cochlear implants and waited 8–10 weeks to let fibrosis form. They then removed the cochleae, fixed and decalcified the bone, and imaged the samples with a spectral-domain OCT microscope, capturing volumes of 400×400×1024 voxels. From these volumes, OCT B-scans were Z-resliced into en face images of the left (implanted) cochlea, with 1024 slices per volume.

Three experienced segmenters at Oregon Health & Science University School of Medicine manually labelled a small number of slices per volume using Hasty.ai tools, assigning three classes: "ST/Free Space", "CI/Track", and "Fibrosis". Every image in the dataset has all three annotated. In total there are 173 annotated images, split 80/20 by stratified random sampling into 139 training and 34 test images. Because one volume (OCTV1L) contributes 77 of the 173 slices, the dataset is imbalanced both by volume and by pixel count per class.

The proposed model is a UNET variant. UNET is an encoder-decoder network: the encoder progressively shrinks feature maps while doubling channel counts to capture context, and the decoder reverses this while skip connections restore fine spatial detail. The authors make it deeper (seven encoder and decoder blocks, 37 Conv2D layers of 3×3 kernels), start with 8 filters, replace Batch Normalization with Group Normalization (n_g = 2) to suit their small batch size of 2, add Dropout (0.1) only in the decoder, and train on 1024×1024 pseudo-RGB inputs upsampled from the original 400×400 images. Weights are initialized with He Normal initialization, and the final layer produces four likelihood maps, reduced to a segmentation by pixel-wise argmax.

Training used an Adam optimizer at learning rate 0.0001, batch size 2, and 100 epochs for 2D-OCT-UNET, UEfficientNet, VGG16-UNET, and MST-DeepLabv3+. A multiclass Dice loss was used for the first three (Dice loss is defined as 1 minus the Dice Score Coefficient, with a smoothing term γ = 1), while MST-DeepLabv3+ used categorical crossentropy. SegFormer was fine-tuned for 200 epochs with Adam at learning rate 0.001 and channel-wise standardization, and SAM was fine-tuned in prompt-free auto-mode with adapters at 1024×1024 image size for a maximum of 200 epochs, stopping early after 78 epochs. No data augmentation and no preprocessing were applied. All work ran on a desktop with Ubuntu 18.04.3 LTS, an Intel Core i9-9900K CPU, an Nvidia GeForce RTX 2080 Ti GPU, and 62GB RAM, using Python 3.9.12, TensorFlow 2.9.1, and Keras 2.9.0, with SAM implemented in PyTorch 2.0.1 (cu117).

Why This Matters

The paper offers an objective, high-resolution, repeatable way to measure fibrotic burden in implanted cochleae — a measurement that is otherwise slow and subject to human variability. Because fibrosis is one candidate mechanism behind the residual hearing loss that affects up to 50% of EAS patients, a reliable quantification tool could sharpen studies of that pathway.

Real-world applications:

• Preclinical implant research. Objectively comparing how much fibrosis different implant designs or surgical techniques provoke in rodent models.
• Assessment across imaging modalities. The authors suggest the tool can take low-resolution ground-truth segmentation as an input, potentially extending to other modalities such as micro computed tomography.
• Human temporal bone studies. The paper speculates it would be of great interest to test 2D-OCT-UNET's utility for human temporal bone imaging, though it notes the caveats around anatomical variability and the implant's different role in humans.
• Accelerating annotation-heavy workflows. By training on sparse human segmentation and applying it to all sections down to the point where signal-to-noise ratio degrades, the model extends labels beyond where human segmenters were confident.

Industry relevance: developers of cochlear implant hardware and hybrid electroacoustic devices have a stake in understanding and minimizing implant-induced tissue response. The work is also relevant to the medical imaging community as a case study in applying segmentation models to a very small, highly imbalanced dataset with noisy labels. The dataset and code are publicly available at https://github.com/juliadietlmeier/CF-OCT-segmentation.

Future Directions

• Growing the dataset and generalizing across volumes. The dataset currently covers 5 animals, and 2D-OCT-UNET's efficacy is limited to within-volume training. The authors plan to expand the Cochlear OCT dataset and pursue cross-volume generalization.

• Reducing ground-truth noise. Future annotations could be performed by two different segmenters and averaged, or noisy labelling could be turned into an opportunity to train the model to control for it more actively.

• Training on volumes without fibrosis. The authors state that including non-fibrotic volumes would introduce greater dataset imbalance but is a necessary next step in model optimization.

• Testing advanced loss functions and other species. The paper flags incorporating imbalance-aware losses such as Unified Focal Loss as future work, and raises the open question of how 2D-OCT-UNET would perform on imaging from other species, including humans.

Target Audience

This paper is most useful to biomedical engineering and computer vision researchers working on medical image segmentation, particularly those dealing with small, severely class-imbalanced datasets and noisy labels. It also serves auditory neuroscience and cochlear implant researchers who need objective ways to measure intracochlear fibrosis, and machine learning practitioners interested in how a relatively modest UNET configuration compares against foundation models and Transformer-based architectures in a low-data medical setting. Readers should have prior familiarity with encoder-decoder segmentation architectures and standard segmentation metrics to follow the benchmark comparisons.

Authors’ abstract

Objective: Cochlear implants (CIs) are bionic prostheses that restores hearing via electrical stimulation of the auditory nerve. Hybrid CIs, which use electroacoustic stimulation (EAS), combine residual low-frequency acoustic hearing with CI electrical stimulation. Intracochlear fibrosis, which forms in response to the presence of the implant, may impede residual hearing function and gradually reduce the efficacy of EAS. It is therefore a translational objective to study the formation of cochlear fibrosis in rodents, with the goal of reducing fibrotic burden and improving outcomes for CI patients. Methods: We generate and annotate a novel dataset of optical coherence tomography (OCT) images from chronically implanted guinea pigs as part of an ongoing study focused on implant induced fibrosis. Objectively assessing fibrotic burden in this model, with high resolution and repeatability, presents an obvious use case for computer vision methods. Results: We present the results of several state-of-the-art semantic segmentation models and compare their efficacy for identifying cochlear fibrosis and other relevant annotations, using a new library of manually segmented OCT images. Conclusions: We find that the best performance is achieved by using a modified version of the well-known UNET architecture (which we term 2D-OCT-UNET) that operates on the upscaled OCT input resolution. Significance: For the first time, we have successfully applied computer vision techniques to an OCT dataset of implanted cochleae with fibrosis. Using this deep learning model, the cochlear fibrotic burden calculation can be reliably carried out as we verify in our experimental section. The dataset and the project code are available at: https://github.com/juliadietlmeier/CF-OCT-segmentation