多区域融合注意力网络模型下的核性白内障分类

Objective: Cataracts are the primary inducement for human blindness and vision impairment. Early intervention and cataract surgery can effectively improve the vision and life quality of cataract patients. Anterior segment optical coherence tomography (AS-OCT) image can capture cataract opacity information through a non-contact, objective, and fast manner. Compared with other ophthalmic images like fundus images, AS-OCT images are capable of capturing the clear nucleus region, which is very significant for nuclear cataract (NC) diagnosis. Clinical studies have identified that a strong opacity correlation relationship and high repeatability between average density value of the nucleus region and NC severity levels in AS-OCT images. Moreover, the clinical works also have suggested that the correlation relationships between different nucleus regions and NC severity levels. These original research works provide the clinical reference for automatic AS-OCT image-based NC classification. However, automatic NC classification based on AS-OCT images has been rarely studied, and there is much improvement room for NC classification performance on AS-OCT images. Method: Motivated by the clinical research of NC, this paper proposes an efficient multi-region fusion attention network (MRA-Net) model by infusing clinical prior knowledge, aiming to classify nuclear cataract severity levels on AS-OCT images accurately. In the MRA-Net, we construct a multi-region fusion attention (MRA) block to fuse feature representation information from different nucleus regions to enhance the overall classification performance, in which we not only adopt the summation operation to fuse different region information but also apply the softmax function to focus on salient channel and suppress redundant channels. In respect of the residual connection can alleviate the gradient vanishing issue, the MRA block is plugged into a cluster of Residual-MRA modules to demonstrate MRA-Net. Moreover, we also test the impacts of two different dataset splitting methods on NC classification results: participant-based splitting method and eye-based splitting method, which is easily ignored by previous works. In the training, this paper resizes the original AS-OCT images into 224 × 224 pixels as the network inputs and set batch size to 16. Stochastic gradient descent (SGD) optimizer is used as the optimizer with default settings and we set training epochs to 100. Result: Our research analysis demonstrates that the proposed MRA-Net achieves 87.78% accuracy and obtains 1% improvement than squeeze and excitation network (SENet) based on a clinical AS-OCT image dataset. We also conduct comparable experiments to verify that the summation operation works better the concatenation on the MRA block by using ResNet as the backbone network. The results of two dataset splitting methods also that ten classification methods like MRA-Net and SENet obtain better classification results on the eye-based dataset than the participant-based dataset, e.g., the highest improvements on F1 and Kappa are 4.03% and 8% correspondingly. Conclusion: Our MRA-Net considers the difference of feature distribution in different regions in a feature map and incorporates the clinical priors into network architecture design. MRA-Net obtains surpassing classification performance and outperforms advanced methods. The classification results of two dataset splitting methods on AS-OCT image dataset also indicated that given the similar nuclear cataract severity in the two eyes of the same participant. Thus, the AS-OCT image dataset is suggested to be split based on the participant level rather than the eye level, which ensures that each participant falls into the same training or testing datasets. Overall, our MRA-Net has the potential as a computer-aided diagnosis tool to assist clinicians in diagnosing cataract.