TTDCapsNet: Tri Texton-Dense Capsule Network for complex and medical image recognition.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Convolutional Neural Networks (CNNs) excel at hierarchical feature learning but require vast datasets.
• Capsule Networks (CapsNets) perform well with limited data but struggle with complex image recognition.

Purpose of the Study:

• To introduce a novel Capsule Network architecture, Tri Texton-Dense CapsNet (TTDCapsNet), for improved complex and medical image classification.
• To address the limitations of existing CNNs and CapsNets in handling complex visual data.

Main Methods:

• Developed TTDCapsNet, a hierarchical architecture comprising three Texton-Dense CapsNet (TDCapsNet) blocks.
• Each TDCapsNet integrates a texton detection layer, an eight-layered dense convolution block, and Primary Capsule (PC) and Class Capsule (CC) layers.
• Employed a routing algorithm to combine feature maps from multiple PCs and CC layers for enhanced classification.

Main Results:

• Achieved high validation accuracies: 94.90% on fashion-MNIST, 89.09% on CIFAR-10, 95.01% on Breast Cancer, and 97.71% on Brain Tumor datasets.
• Demonstrated superior performance compared to baseline models and competitive results against state-of-the-art CapsNet models.
• Confirmed the effectiveness of the routing algorithm and the hierarchical structure in improving classification outcomes.

Conclusions:

• The proposed TTDCapsNet model is viable for complex real-world image classification tasks.
• This architecture shows significant potential as an intelligent system for aiding oncologists in disease diagnosis and treatment planning.
• The study highlights the advancement of Capsule Networks in tackling challenging visual recognition problems.