Abstract:Early diagnosis of patients' disease is crucial since it helps doctors and patients devise a treatment plan. Therefore, recognizing medical images using Artificial intelligence-based deep learning techniques has recently increased. Capsule Network (CapsNet) has promising methods in visual tasks due to its ability to keep a high relationship of spatial information compared to convolutional neural networks (CNNs). However, CapsNet faces a critical problem with a complex image background that limits its performance. The traditional CapsNet adopts a standalone convolution (SC) as a feature extractor, Softmax function for normalization of coupling coefficient, and dynamic routing procedure to allow active capsules to perform predictions leading to activation of high-level capsules. The SC is not an effective feature extractor, and SoftMax impedes capsules from distributing optimal coupling coefficient during routing. This paper proposes a CapsNet architecture called SqueezeCapsNet that integrates SqueezeNet and CapsNet to achieve effective feature extraction and fewer parameters. A new squash function named parametric squash function (PSF) was proposed to reduce non-informative capsules and promote discriminative capsules. To the best of our knowledge in literature, we are the first to integrate SqueezeNet into CapsNet. We evaluate our framework on two medical image datasets; Brain tumor and Lung & Colon cancer datasets. Additionally, datasets with varied backgrounds; MNIST, fashion-MNIST, CIFAR-10 were used to evaluate the robustness and generalizability of the model. The SqueezeCapsNet produces 94.85%, 99.76%, 99.87 % , 93.49 % , and 82.45 % on Brain tumor, Lung & Colon Cancer, MNIST, fashion-MNIST, and CIFAR-10 datasets, respectively. Experimental results show that the proposed architecture's compression techniques significantly provide fewer parameters while enhancing stability and accuracy across all the evaluation metrics. Our results show that our method improves CapsNet and can be adopted as a computer-aided diagnostic method to support the diagnosis of medical image tasks.

Next-gen image enhancement: CapsNet-driven auto-encoder model in single image super resolution

Single Image Super-Resolution Based on Capsule Neural Networks

DeepCaps: Going Deeper With Capsule Networks

RS-CapsNet: an Advanced Capsule Network.

Feature Amplification Capsule Network for Complex Images.

Fast and Accurate Single Image Super-Resolution Via an Energy-Aware Improved Deep Residual Network.

Deep Networks for Image and Video Super-Resolution

CVANet: Cascaded visual attention network for single image super-resolution

Dual Directed Capsule Network for Very Low Resolution Image Recognition

Dual-path attention network for single image super-resolution

SqueezeCapsNet: enhancing capsule networks with squeezenet for holistic medical and complex images

Image Superresolution using Scale-Recurrent Dense Network

CA-ESRGAN: Super-Resolution Image Synthesis Using Channel Attention-Based ESRGAN

Single Image Super-Resolution Via Residual Neuron Attention Networks.

A Single Super-Resolution Method Via Deep Cascade Network

EPNet: An Efficient Pyramid Network for Enhanced Single-Image Super-Resolution with Reduced Computational Requirements

RSAMSR: A deep neural network based on residual self-encoding and attention mechanism for image super-resolution

Image Superresolution Via Dense Discriminative Network.

Image Super-Resolution Reconstruction Model Based on Multi-Feature Fusion

Context-Aware Residual Network With Promotion Gates For Single Image Super-Resolution

Multi-Path Deep CNN with Residual Inception Network for Single Image Super-Resolution