Core Concepts
Integrating Capsule Networks with Graph Neural Networks improves skin cancer classification accuracy significantly.
Abstract
In the realm of skin lesion image classification, challenges arise from spatial and semantic features. Previous methods struggle due to imbalanced datasets. This study introduces a novel approach by combining Graph Neural Networks (GNNs) with Capsule Networks to enhance classification performance. The hybrid model was applied to the MNIST:HAM10000 dataset, achieving an accuracy improvement of 95.52%. The research focuses on overcoming challenges in skin lesion classification, contributing to image-based diagnosis in dermatology. GNNs offer advanced mechanisms for capturing complex patterns beyond traditional CNNs. Capsule Networks recognize spatial hierarchies within images effectively.
Stats
After 75 epochs of training, the model achieved an accuracy improvement reaching 89.23% and 95.52%.
Established benchmarks such as GoogLeNet (83.94%), InceptionV3 (86.82%), MobileNet V3 (89.87%), EfficientNet-7B (92.07%), ResNet18 (92.22%), ResNet34 (91.90%), ViT-Base (73.70%), and IRv2-SA (93.47%) were surpassed on the same dataset.
Quotes
"Our research focuses on evaluating and enhancing the Tiny Pyramid Vision GNN architecture by incorporating it with a Capsule Network."
"GNNs offer an advanced mechanism for capturing complex patterns beyond the capabilities of traditional CNNs."
"Capsule Networks provide superior recognition of spatial hierarchies within images."