A Novel Architecture of Depthwise Separable CNN and Multi-Level Pooling for Detection and Classification of Myopic Maculopathy

Abstract

Myopic maculopathy (MM), also known as myopic macular degeneration, is the most serious, irreversible, vision-threatening complication and the leading cause of visual impairment and blindness. Numerous research studies demonstrate that the convolutional neural network (CNN) outperforms many applications. Current CNN designs employ a variety of techniques, such as fixed convolutional kernels, the absolute value layer, data augmentation, and domain knowledge, to enhance performance. However, some network structure designing hasn't received much attention yet. The intricacy of the MM categorization and definition system makes it challenging to employ deep learning (DL) technology in the diagnosis of pathologic myopia lesions. To increase the detection precision of MM's spatial domain, the proposed work first concentrates on creating a novel CNN network structure then improve the convolution kernels in the preprocessing layer. The number of parameters is decreased, and the characteristic of a small local region is modeled using the smaller convolution kernels. Next channel correlation of the residuals with separable convolutions is employed to compress the image features. Then, the local features using the spatial pyramid pooling (SPP) technique is combined, which improves the features' capacity to be represented by multi-level pooling. The use of data augmentation is the final step in enhancing network performance. Compress the residuals in this paper to make use of the channel correlation. The accuracy achieved by the model was 95%, F1-score of 96.5% and AUC of 0.92 on augmented MM-PALM dataset. The paper concludes by conducting a comparative study of various deep-learning architectures. The findings highlight that the hybrid CNN with SPP and XgBoost (Depthwise-XgBoost) architecture is the ideal deep learning classification model for automated detection of four stages of MM.