Source-linked AI summary
Towards Investigating Residual Hearing Loss: Quantification of Fibrosis in a Novel Cochlear OCT Dataset
Julia Dietlmeier, Benjamin Greenberg, Wenxuan He, Teresa Wilson, Rubing Xing, Jordan Hill, Adrienne Fettig, Madeline Otto, Teyhana Rounsavill, Lina A. J. Reiss, Jingang Yi, Noel E. O'Connor, George W. S. Burwood
TL;DR
Residual hearing loss after cochlear implantation remains multifactorial, and the role of intracochlear fibrosis is still under investigation. The paper creates an annotated OCT dataset from implanted guinea pigs and benchmarks semantic segmentation models. Modified 2D-OCT-UNET performs best, achieving multiclass DSC of 0.8874, while the dataset and model provide a proof of concept within a small, within-volume setting.
Problem
Residual hearing loss affects up to 50% of EAS patients, while the relationship between cochlear fibrosis and residual hearing mechanics remains under investigation.
Method
The authors curate and manually annotate a cochlear OCT dataset from five chronically implanted guinea pigs and compare state-of-the-art semantic segmentation models.
Results
Multiclass DSC of 0.8874 was achieved by 2D-OCT-UNET, the best-performing model among six evaluated CNN- and vision-Transformer-based models.
Takeaways & Limitations
The work demonstrates a computer-vision proof of concept for quantifying cochlear fibrotic burden in implanted-cochlea OCT images.
Takeaways & Limitations
The dataset contains only five animals, and 2D-OCT-UNET efficacy is currently limited to within-volume training.
Abstract
from arXiv · showhide
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
I. BACKGROUND
Residual hearing loss after cochlear implantation is common but multifactorial, motivating study of fibrosis with OCT and computer vision. The paper introduces an annotated OCT dataset and evaluates segmentation for cochlear fibrotic burden.
- Motivation: Up to 50% of EAS patients lose residual hearing within months to years after cochlear implantation.Reported causes are unknown and multifactorial, including neuronal and hair-cell loss, surgical trauma, and immune response.
- Motivation: Fibrotic tissue and eventual ossification around the CI array are hypothesized to disrupt cochlear mechanics.Modeling and acute rodent studies suggest that CI or fibrotic burden could interfere with cochlear mechanical function.
- Motivation: OCT offers high-resolution, label-free imaging that can support objective and rapid estimates of cochlear fibrotic burden.The study positions OCT datasets as a use case for machine-learning segmentation algorithms.
- Dataset: The dataset contains 173 manually annotated images from five chronically implanted guinea-pig volumes, with Fibrosis, ST/Free Space, CI/Track, and Background classes.The images are split into 139 training and 34 testing samples, with substantial class and slice imbalance.
- Dataset: The OCT dataset is challenging because the three target objects vary in shape, size, and appearance and can closely abut normal tissue.All classes partially cover black pixels, complicating semantic segmentation.
III. ARCHITECTURES FOR SEMANTIC SEGMENTATION OF COCHLEAR OCT DATA
The paper frames semantic segmentation as pixel-wise categorization and compares CNN, Transformer, and foundation-model architectures for multiclass cochlear OCT segmentation.
- CNNs: CNNs support image segmentation through convolutional, activation, and pooling layers.The passage describes convolution as a learnable-kernel dot product over a local receptive field.
- Transformers and foundation models: Vision Transformers process image patches with attention and can outperform CNNs on downstream tasks when pretrained on large datasets.Foundation models extend this paradigm through large-scale pretraining and downstream adaptation.
- Task: Semantic segmentation categorizes each image pixel into an object class, here using Fibrosis, ST/Free Space, CI/Track, and Background.The task segments three non-overlapping cochlear objects plus background.
- Compared architectures: The evaluated model families include UNET-based CNNs, Segment Anything Model, SegFormer, and MST-DeepLabv3+.These architectures span conventional convolutional, Transformer, foundation-model, and hybrid designs.
A. Proposed 2D-OCT-UNET configuration
The proposed 2D-OCT-UNET adapts the fully supervised UNET for small-data cochlear OCT segmentation through deep encoder-decoder processing, skip connections, and high-resolution inputs.
- UNET foundation: UNET uses a symmetric encoder-decoder with skip connections to share multiscale information during image segmentation.Encoder blocks extract context while decoder blocks upsample and recover spatial information.
- Proposed configuration: 2D-OCT-UNET contains seven encoder-decoder blocks and uses Group Normalization instead of Batch Normalization.Group Normalization is selected for the small batch size of 2.
- Architecture depth: The network has 37 Conv2D layers and increases filter counts after each MaxPool2D layer to model long-range feature dependencies.The authors associate this depth with a mechanism commonly attributed to Transformer self-attention.
- Input and training: The model uses 1024 × 1024 pseudo-RGB inputs, obtained by resizing the original 400 × 400 OCT images, and is trained from scratch.No transfer learning is used in this configuration.
- Output: The model outputs four likelihood maps, and pixel-wise argmax produces the final multiclass segmentation.The four outputs correspond to the number of classes, including background.
C. UEfficientNet
UEfficientNet combines a pretrained EfficientNet encoder with UNET and residual decoder learning, while the compared alternatives use Transformer, attention, and lightweight-backbone designs.
- UEfficientNet: UEfficientNet combines an ImageNet-pretrained EfficientNet-B4 encoder with a UNET decoder using residual learning.Its decoder differs from VGG16-UNET through residual decoder blocks.
- UEfficientNet: EfficientNet balances network depth, breadth, and resolution through compound coefficient optimization for improved accuracy and resource use.UEfficientNet uses 256 × 256 input images.
- SegFormer: SegFormer combines a hierarchical Transformer encoder with lightweight MLP decoders and local-global attention.The encoder produces multiscale features through self-attention.
- MST-DeepLabv3+: MST-DeepLabv3+ uses a MobileNetV2 backbone, SENet channel attention, and ImageNet transfer learning.The design targets improved results with fewer training parameters.
F. Segment Anything Model (SAM)
SAM is a vision Transformer foundation model fine-tuned for prompt-free, multiclass cochlear OCT segmentation. The study uses adapter-based fine-tuning to enable a fair comparison on its new dataset.
- SAM is a vision Transformer-based image segmentation foundation model trained on over 11 million natural images and one billion segmentation masks.
- The model is fine-tuned because the study requires prompt-free, multiclass segmentation on cochlear OCT images.
- Adapter blocks are added to the mask decoder while using the ViT-B image encoding architecture.
- The imaging study uses chronically implanted guinea pigs with an 8–10 week post-surgery latency period for intracochlear fibrosis formation.
C. OCT imaging
Cochlear OCT volumes are acquired with spectral-domain OCT at high spatial sampling and converted into grayscale image sequences for manual segmentation.
- A 4mm^2 × 3.5mm volume with 400×400×1024 voxels is acquired in 3D mode at a 10 kHz sampling rate.
- Imaging uses a Telesto III spectral-domain OCT microscope with a 1300 nm central wavelength and 5× objective lens.
- The basal turn is positioned in the coherence gate, and all three scalae of the basal turn are imaged.
- OCT spectra are converted to 8-bit grayscale image sequences in ImageJ before manual segmentation in Hasty.ai.
D. Implementation details
The models are trained with Adam-based optimization and model-specific losses, while preprocessing and training schedules vary across architectures.
- 2D-OCT-UNET, UEfficientNet, VGG16-UNET and MST-DeepLabv3+ use Adam optimization with a learning rate of 0.0001, batch size 2 and 100 epochs.
- The CNN models use multiclass Dice loss except MST-DeepLabv3+, which uses categorical crossentropy.
- Dice loss equals 1 − DSC, where DSC measures overlap between predicted and ground-truth segmentations from 0 to 1.
- The multiclass Dice formulation uses C for classes, n for class samples, and γ for smoothing, generally set to 1.
- No data augmentation or preprocessing is applied for the CNN training procedure.
- SegFormer is fine-tuned for 200 epochs with Adam at learning rate 0.001, channel-wise standardization and 512 × 512 resizing.
- SAM uses prompt-free adapter fine-tuning with batch size 2 and 1024 × 1024 images, stopping early after 78 epochs.
E. Experimental results
Six semantic segmentation models are benchmarked using standard metrics, and 2D-OCT-UNET provides the strongest overall performance while fibrosis-burden estimates show volume-specific accuracy differences and no significant absolute-error correlation with fibrosis amount.
- 2D-OCT-UNET outperforms the other models in four of five metrics, while SAM adapters improve performance but remain below 2D-OCT-UNET.
- SegFormer trails 2D-OCT-UNET by 0.0794 on DSC and produces pixelated edge artefacts in qualitative segmentation maps.
- The authors reserve testing advanced loss functions for future work because imbalanced semantic segmentation remains incompletely understood.
- Fibrotic burden is computed as the fibrosis area divided by fibrosis plus ST/Free Space area, multiplied by 100.
- AI-computed fibrosis curves are compared with sparser manually segmented areas for volumes OCTV1L and OCTV7L.
- Prediction accuracy is lower for OCTV1L than OCTV7L, with signed and absolute errors defined relative to ground-truth fibrosis values.
- Absolute error has no significant correlation with fibrosis amount for OCTV1L (ρ = −0.1088, p = 0.3464) or OCTV7L (ρ = 0.0374, p = 0.8723).
V. DISCUSSION
The study applies 2D-OCT-UNET to quantify cochlear fibrotic burden, while identifying annotation noise, limited data, and restricted within-volume training as important boundaries. Its broader applicability to other species and imaging modalities remains uncertain.
- 2D-OCT-UNET was developed and applied to compute cochlear fibrotic burden for investigating residual hearing loss.
- Annotation noise arose where the basilar membrane–fibrosis interface was difficult to identify, although the model often excluded existing tissue from fibrosis labels.
- Averaging two or more human measurements is proposed to reduce ground-truth error from noisy labeling.
- Fibrotic-burden computation was less accurate for OCTV1L, which the authors associate with noisier labels and more imbalanced semantic annotations.
- The OCTV7L ground-truth outlier was tolerated and ignored by 2D-OCT-UNET, suggesting robustness to certain levels of annotation noise.
- The dataset contains five animals, and 2D-OCT-UNET efficacy is currently limited to within-volume training for guinea-pig cochlear OCT scans.
- Performance on imaging from other species, including humans, remains unclear because of increased anatomical variability and differing clinical implant contexts.
VI. CONCLUSION
The paper introduces an annotated cochlear OCT dataset and semantic-segmentation task for quantifying intracochlear fibrosis, with 2D-OCT-UNET achieving the strongest reported performance. The work supports future investigation of relationships between cochlear fibrosis and residual hearing loss, while noting that training data currently exclude volumes without fibrosis.
- The study curates an annotated cochlear OCT dataset from chronically implanted guinea pigs exhibiting intracochlear fibrosis.
- It introduces cochlear semantic segmentation to study and quantify cochlear fibrotic burden.
- An annotation outlier in the OCTV7L ground truth was tolerated and ignored by 2D-OCT-UNET, although its effect would have produced p-value 0.0050 in formal error analysis.
- The task has highly variable object shapes, sizes, and appearances, and 2D-OCT-UNET was trained only on implanted cochleae.Training on volumes without fibrosis is identified as a necessary next step, despite increasing dataset imbalance.
- 2D-OCT-UNET achieved the best performance among six benchmarked CNN- and vision Transformer-based models, reaching multiclass DSC of 0.8874.Qualitative results also verified high-quality semantic segmentation.
- The findings could advance future studies examining the relationship between cochlear fibrosis and residual hearing loss.
COMPLIANCE WITH ETHICAL GUIDELINES
All procedures complied with OHSU IACUC protocols and received permission from the local ethics committee.
- The study followed OHSU IACUC protocols and local ethics committee approval.Protocol IP00000453 was approved on 4/6/2022; IP00001278 was approved on 12/29/2020 and renewed on 12/06/2023.