Abstract
Glaucoma can lead to cellular damage within retinal layers, ultimately leading to vision loss and potential blindness. Optical coherence tomography (OCT) imaging has emerged as a valuable tool for diagnosing and tracking glaucoma, offering detailed insights into the deep retinal layers. However, due to the intrinsic limitation of localized focus in OCT imaging, models that rely solely on OCT images captured by a single scanning pattern hold a high risk of losing valuable features. To address this challenge, we propose a novel model for glaucoma detection, employing OCT images captured by four scanning patterns from three fundus regions. Additionally, it’s crucial to note that glaucoma is irreversible, necessitating continuous tracking of its progression is essential following diagnosis. Therefore, we propose a cost-efficient and glaucoma-specific segmentation model for retinal layers in OCT images by exploiting shared features from low-cost normal OCT images. Subsequently, to adapt various clinic scenarios and color fundus images, we propose a novel adaptable multi-scenario and glaucoma-specific model for optic disc (OD) and optic cup (OC) segmentation with style contrastive and domain adversarial collaborative learning. Finally, due to limited resources in clinical practice, we propose a source- and framework-free model for glaucoma-specific segmentation tasks based on the fully test time adaptative learning and semi-supervised learning, only reusing parameters and modules from a source-trained model, and without requiring target annotations.In the first study, to mitigate the high risk of losing features in other fundus regions and scanning patterns in existing single scan-pattern-based glaucoma detection models, we collect and construct a dataset encompassing comprehensive OCT images captured by four scanning patterns from three fundus regions. Leveraging this dataset, we propose a novel model that integrates multi-region and multi-scan-pattern OCT images. To effectively integrate features from multiple scanning patterns and multiple fundus regions, we propose two novel feature fusion modules. The first, termed as the attention multiple scan-pattern feature fusion module, aims at fusing features from multiple scan-patterns within a given fundus region. The second, named by the attention multiple regions feature fusion module, is designed to integrate features from multiple fundus regions, concurrently assigning suitable weights according to the importance of each region, thereby enabling ophthalmologists to prioritize critical fundus regions. Compared with the common average fusion strategy, our proposed feature fusion modules exhibit enhanced adaptability across different scenarios, yielding enhanced detection performance. Furthermore, we conduct an in-depth investigation to explore the relationship between the single scan-pattern and single region-based models and their relevance to glaucoma.
In the second study, we propose a cost-efficient and glaucoma-specific segmentation model for retinal layers analysis in OCT images, designed to assist in tracking glaucoma progression in confirmed cases. The proposed model employs knowledge transfer learning to exploit resources from normal OCT images, reducing data collection and annotation expenses, and thereby avoiding the high cost of training a fully supervised model from scratch. The proposed model leverages the inherent similarities between normal and glaucoma OCT images, enabling the extraction of shared features from low-cost normal OCT images. Initially, we exploit shared layout and texture features within the output space at a suitable assessment scale. Subsequently, we exploit high-dimensional shared features across various encoding spaces. We then integrate the shared features in both encoding and output spaces. Moreover, we have collected and constructed a glaucoma OCT dataset with detailed annotations of multiple retinal layers. The results demonstrate that our proposed model exhibits superior performance compared to existing alternatives in terms of segmentation and glaucoma-related metrics. Consequently, the proposed model enables providing accurate retinal layer segmentation results for assisting in tracking glaucoma progression.
In the third study, we propose an adaptable multi-scenario and glaucoma-specific OD/OC segmentation model that combines style contrastive learning with domain adversarial learning, specifically targeting color fundus images: a low-cost imaging modality relevant to glaucoma. Moreover, the proposed model is designed to adapt three distinct scenarios: (1) a dataset containing only annotated normal samples, (2) a dataset containing only annotated glaucoma samples, and (3) a dataset that includes both annotated glaucoma and normal samples. Our evaluation begins by independently employing style contrastive learning or domain adversarial learning. Subsequently, we extend the analysis by integrating both learning strategies across three clinical scenarios. The results demonstrate that our proposed model yields superior performance over the baseline models across all scenarios. Moreover, the proposed model achieves a cost-efficient enhancement in performance by solely increasing the amount of low-cost normal samples. Moreover, extensive domain adaptative experiments on two public datasets further confirm the model’s domain adaptability across different clinical centers.
In the fourth study, to address the limited resources commonly encountered in clinical practices, which renders most cost-effective methods impractical, we propose a glaucoma-specific segmentation model that is source- and framework-free, only reusing the parameters and modules from the source-trained model. Initially, we directly extend the existing SOTA fully test-time adaptative models to our target task based on the similar scenario, demonstrating promising results. Subsequently, motivated by the shared challenge of effectively utilizing unlabeled samples in semi-supervised learning, we integrate weak-to-strong consistency perturbations into our proposed model, leading to enhanced performance. Moreover, to align with our target task, based on prior knowledge, we incorporate retinal layers flattening and elastic deformation perturbations into weak and strong branches, respectively, which enhances performance. Additionally, we propose a novel class- and sample-wise threshold method, effectively removing unreliable pseudo-labels. Furthermore, we propose a novel uncertainty reduction strategy by incorporating uncertainty from each single branch prediction and the ensembled averaged prediction, thereby preserving the diversity of various samples. Moreover, the proposed model effectively mitigates domain shifts arising from samples captured by different clinical centers. The experimental results demonstrate that the proposed model exhibits promising performance on both OCT and color fundus images.
In summary, we propose a novel glaucoma detection model by employing multi-scan-pattern and multi-regions OCT images, which mitigates the risk of losing features associated with relying solely on a single scanning pattern or a single region. Furthermore, for the patients with confirmed glaucoma, we propose three glaucoma-specific segmentation models for adapting different clinical scenarios and various imaging, effectively assisting in tracking glaucoma progression and preventing further vision loss.
| Date of Award | 27 Dec 2024 |
|---|---|
| Original language | English |
| Awarding Institution |
|
| Supervisor | Jicong ZHANG (External Supervisor), Shiqi WANG (Supervisor) & Shuaicheng LI (Supervisor) |