Perturbed Composite Attention Model for Macular Optical Coherence Tomography Image Classification

Abstract

In this article, we propose a deep architecture stemming from a perturbed composite attention mechanism with the following two novel attention modules: Multilevel perturbed spatial attention (MPSA) and multidimension attention (MDA) for macular optical coherence tomography (OCT) image (scan) classification. MPSA is designed by adding positive perturbations to the attention layers, thereby amplifying both the salient regions of input images and discriminative features obtained from intermediate layers of the network. On the other hand, the MDA encodes the normalized interdependency of spatial information among various channels of the extracted feature maps. The perturbed composite attention mechanism enables the new architecture to automatically extract relevant diagnostic features at different levels of feature representation resulting in the superior classification of macular diseases such as age-related macular degeneration (AMD), diabetic macular edema (DME), and choroidal neovascularization (CNV). The proposed end-to-end trainable architecture does not require preprocessing steps, such as region of interest extraction, denoising, and retinal flattening, making the network more robust and fully automatic. Experimental results on three macular OCT datasets and ablation studies show that our proposed network outperforms the current state-of-the-art methodologies.