Can Glaucoma Suspect Data Help to Improve the Performance of Glaucoma Diagnosis?

Purpose: The presence of imbalanced datasets in medical applications can negatively affect deep learning methods. This study aims to investigate how the performance of convolutional neural networks (CNNs) for glaucoma diagnosis can be improved by addressing imbalanced learning issues through utilizing glaucoma suspect samples, which are often excluded from studies because they are a mixture of healthy and preperimetric glaucomatous eyes, in a semi-supervised learning approach. Methods: A baseline 3D CNN was developed and trained on a real-world glaucoma dataset, which is naturally imbalanced (like many other real-world medical datasets). Then, three methods, including reweighting samples, data resampling to form balanced batches, and semi-supervised learning on glaucoma suspect data were applied to practically assess their impacts on the performances of the trained methods. Results: The proposed method achieved a mean accuracy of 95.24%, an F1 score of 97.42%, and an area under the curve of receiver operating characteristic (AUC ROC) of 95.64%, whereas the corresponding results for the traditional supervised training using weighted cross-entropy loss were 92.88%, 96.12%, and 92.72%, respectively. The obtained results show statistically significant improvements in all metrics. Conclusions: Exploiting glaucoma suspect eyes in a semi-supervised learning method coupled with resampling can improve glaucoma diagnosis performance by mitigating imbalanced learning issues. Translational Relevance: Clinical imbalanced datasets may negatively affect medical applications of deep learning. Utilizing data with uncertain diagnosis, such as glaucoma suspects, through a combination of semi-supervised learning and class-imbalanced learning strategies can partially address the problems of having limited data and learning on imbalanced datasets.

Mixed multiscale BM4D for three-dimensional optical coherence tomography denoising.

A multiscale extension for the well-known block matching and 4D filtering (BM4D) method is proposed by analyzing and extending the wavelet subbands denoising method in such a way that the proposed method avoids directly denoising detail subbands, which considerably simplifies the computations and makes the multiscale processing feasible in 3D. To this end, we first derive the multiscale construction method in 2D and propose multiscale extensions for three 2D natural image denoising methods. Then, the derivation is extended to 3D by proposing mixed multiscale BM4D (mmBM4D) for optical coherence tomography (OCT) image denoising. We tested mmBM4D on three public OCT datasets captured by various imaging devices. The experiments revealed that mmBM4D significantly outperforms its original counterpart and performs on par with the state-of-the-art OCT denoising methods. In terms of peak-signal-to-noise-ratio (PSNR), mmBM4D surpasses the original BM4D by more than 0.68 decibels over the first dataset. In the second and third datasets, significant improvements in the mean to standard deviation ratio, contrast to noise ratio, and equivalent number of looks were achieved. Furthermore, on the downstream task of retinal layer segmentation, the layer quality preservation of the compared OCT denoising methods is evaluated.

Automated summarisation of SDOCT volumes using deep learning: Transfer learning vs de novo trained networks.

Spectral-domain optical coherence tomography (SDOCT) is a non-invasive imaging modality that generates high-resolution volumetric images. This modality finds widespread usage in ophthalmology for the diagnosis and management of various ocular conditions. The volumes generated can contain 200 or more B-scans. Manual inspection of such large quantity of scans is time consuming and error prone in most clinical settings. Here, we present a method for the generation of visual summaries of SDOCT volumes, wherein a small set of B-scans that highlight the most clinically relevant features in a volume are extracted. The method was trained and evaluated on data acquired from age-related macular degeneration patients, and "relevance" was defined as the presence of visibly discernible structural abnormalities. The summarisation system consists of a detection module, where relevant B-scans are extracted from the volume, and a set of rules that determines which B-scans are included in the visual summary. Two deep learning approaches are presented and compared for the classification of B-scans-transfer learning and de novo learning. Both approaches performed comparably with AUCs of 0.97 and 0.96, respectively, obtained on an independent test set. The de novo network, however, was 98% smaller than the transfer learning approach, and had a run-time that was also significantly shorter.

Automated segmentation of mouse OCT volumes (ASiMOV): Validation & clinical study of a light damage model.

The use of spectral-domain optical coherence tomography (SD-OCT) is becoming commonplace for the in vivo longitudinal study of murine models of ophthalmic disease. Longitudinal studies, however, generate large quantities of data, the manual analysis of which is very challenging due to the time-consuming nature of generating delineations. Thus, it is of importance that automated algorithms be developed to facilitate accurate and timely analysis of these large datasets. Furthermore, as the models target a variety of diseases, the associated structural changes can also be extremely disparate. For instance, in the light damage (LD) model, which is frequently used to study photoreceptor degeneration, the outer retina appears dramatically different from the normal retina. To address these concerns, we have developed a flexible graph-based algorithm for the automated segmentation of mouse OCT volumes (ASiMOV). This approach incorporates a machine-learning component that can be easily trained for different disease models. To validate ASiMOV, the automated results were compared to manual delineations obtained from three raters on healthy and BALB/cJ mice post LD. It was also used to study a longitudinal LD model, where five control and five LD mice were imaged at four timepoints post LD. The total retinal thickness and the outer retina (comprising the outer nuclear layer, and inner and outer segments of the photoreceptors) were unchanged the day after the LD, but subsequently thinned significantly (p < 0.01). The retinal nerve fiber-ganglion cell complex and the inner plexiform layers, however, remained unchanged for the duration of the study.