Transformations performed for data augmentation.

Age-related macular degeneration (AMD) is an eye disease that leads to the deterioration of the central vision area of the eye and can gradually result in vision loss in elderly individuals. Early identification of this disease can significantly impact patient treatment outcomes. Furthermore, given the increasing elderly population globally, the importance of automated methods for rapidly monitoring at-risk individuals and accurately diagnosing AMD is growing daily. One standard method for diagnosing AMD is using optical coherence tomography (OCT) images as a non-invasive imaging technology. In recent years, numerous deep neural networks have been proposed for the classification of OCT images. Utilizing pre-trained neural networks can speed up model deployment in related tasks without compromising accuracy. However, most previous methods overlook the feasibility of leveraging pre-existing trained networks to search for an optimal architecture for AMD staging on a new target dataset. In this study, our objective was to achieve an optimal architecture in the efficiency-accuracy trade-off for classifying retinal OCT images. To this end, we employed pre-trained medical vision transformer (MedViT) models. MedViT combines convolutional and transformer neural networks, explicitly designed for medical image classification. Our approach involved pre-training two distinct MedViT models on a source dataset with labels identical to those in the target dataset. This pre-training was conducted in a supervised manner. Subsequently, we evaluated the performance of the pre-trained MedViT models for classifying retinal OCT images from the target Noor Eye Hospital (NEH) dataset into the normal, drusen, and choroidal neovascularization (CNV) classes in zero-shot settings and through five-fold cross-validation. Then, we proposed a stitching approach to search for an optimal model from two MedViT family models. The proposed stitching method is an efficient architecture search algorithm known as stitchable neural networks. Stitchable neural networks create a candidate model in search space for each pair of stitchable layers by inserting a linear layer between them. A pair of stitchable layers consists of layers, each selected from one input model. While stitchable neural networks had previously been tested on more extensive and general datasets, this study demonstrated that stitching networks could also be helpful in smaller medical datasets. The results of this approach indicate that when pre-trained models were available for OCT images from another dataset, it was possible to achieve a model in 100 epochs with an accuracy of over 94.9% in classifying images from the NEH dataset. The results of this study demonstrate the efficacy of stitchable neural networks as a fine-tuning method for OCT image classification. This approach not only leads to higher accuracy but also considers architecture optimization at a reasonable computational cost.

年龄相关性黄斑变性（Age-related macular degeneration, AMD）是一种可导致眼部中央视觉区域退化，并逐渐造成老年人群视力丧失的眼部疾病。早期识别该疾病对患者的治疗结局具有显著影响。此外，随着全球老年人口规模持续扩大，用于快速筛查高危人群并精准诊断AMD的自动化方法的重要性与日俱增。诊断AMD的标准手段之一，是采用光学相干断层扫描（optical coherence tomography, OCT）成像技术——这是一种非侵入式成像方式。近年来，诸多用于OCT图像分类任务的深度神经网络被相继提出。借助预训练神经网络，可在不损失模型准确率的前提下，加快相关任务中的模型部署速度。但绝大多数既往研究均忽略了一个可行性方向：利用已预训练完成的网络，在全新的目标数据集上搜索适配AMD分期任务的最优模型架构。本研究的目标，是在视网膜OCT图像分类任务中实现效率与准确率的最优权衡。为此，我们采用了预训练医学视觉Transformer（Medical Vision Transformer, MedViT）模型。MedViT融合了卷积神经网络与Transformer神经网络，专为医学图像分类任务量身打造。我们的具体方案为：先在标签集与目标数据集完全一致的源数据集上，以监督学习方式预训练两款不同的MedViT模型。随后，我们分别在零样本设置下，以及通过五折交叉验证两种方式，评估了预训练MedViT模型对目标数据集——努尔眼科医院（Noor Eye Hospital, NEH）数据集的视网膜OCT图像的分类性能，该数据集涵盖正常、玻璃膜疣以及脉络膜新生血管（choroidal neovascularization, CNV）三个类别。此后，我们提出了一种拼接方法，用于从两款MedViT系列模型中搜索最优模型。所提拼接方法属于一类高效的架构搜索算法，即可拼接神经网络（stitchable neural networks）。可拼接神经网络通过在每一对可拼接层之间插入线性层，在搜索空间中为每一组可拼接层对生成候选模型；其中每一对可拼接层分别选自两个输入模型中的某一层。尽管此前可拼接神经网络已在规模更大的通用数据集上完成验证，但本研究证明，该方法同样可应用于规模较小的医学数据集。本方法的实验结果显示：当存在其他数据集上预训练的OCT图像模型时，仅需100个训练轮次，即可在NEH数据集的图像分类任务中获得超过94.9%的准确率。本研究结果证实了可拼接神经网络作为OCT图像分类微调方法的有效性。该方法不仅能够实现更高的分类准确率，同时还能在合理的计算成本下兼顾模型架构优化。