A multi-stage fusion deep learning framework merging local patterns with attention-driven contextual dependencies for...

Area of Science:

• Artificial Intelligence in Medicine
• Deep Learning for Medical Imaging
• Oncology Diagnostics

Background:

• Early cancer diagnosis is crucial but challenged by expert shortages and complex disease patterns.
• Deep learning models offer potential for faster, more accurate disease detection.
• Existing methods face limitations in extracting comprehensive features for complex medical images.

Purpose of the Study:

• To propose novel deep learning models for detecting brain tumors, lung cancer, and skin cancer.
• To evaluate the efficacy of the PADBSRNet architecture and a PADBSRNet-Vision Transformer (ViT) hybrid.
• To enhance the speed and accuracy of cancer diagnosis through advanced AI.

Main Methods:

• Development of PADBSRNet, a deep neural network integrating convolutional layers, attention mechanisms, and recurrent networks.
• Creation of a hybrid model combining PADBSRNet with Vision Transformer (ViT) capabilities.
• Experimental validation using diverse medical datasets including brain tumor, lung, and skin cancer images.

Main Results:

• The PADBSRNet model demonstrated high accuracy: 95.24% for brain tumors, 99.55% for lung cancer, and 88.61% for skin cancer.
• The proposed models effectively extracted local-global and contextual features.
• The deep learning approach successfully modeled long-term dependencies crucial for accurate classification.

Conclusions:

• The PADBSRNet model and its hybrid variant show significant promise for accurate and efficient cancer diagnosis.
• Deep learning models can effectively learn intricate patterns in medical images, aiding clinical decision-making.
• These AI-driven approaches offer a viable solution to improve public health outcomes in oncology.