CroReLU: Cross-Crossing Space-Based Visual Activation Function for Lung Cancer Pathology Image Recognition

Area of Science:

• Computational pathology
• Artificial intelligence in oncology
• Deep learning for medical imaging

Background:

• Lung cancer diagnosis relies heavily on pathology, considered the gold standard.
• Increasing patient numbers and image complexity strain pathologist capacity, especially in resource-limited regions.
• Current deep learning models require optimization for accurate pathology image analysis.

Purpose of the Study:

• To introduce CroReLU, a novel plug-and-play visual activation function (AF) for deep learning models in lung cancer pathology.
• To leverage a priori pathological knowledge to improve the diagnostic accuracy of AI models.
• To enhance the capability of deep learning in precision medicine for lung cancer.

Main Methods:

• Developed CroReLU, an activation function with a unique crossover window design for neural networks.
• Integrated CroReLU into the SeNet architecture, creating SeNet_CroReLU.
• Trained and validated SeNet_CroReLU on a dataset of 776 lung cancer pathology images and the LC25000 dataset.

Main Results:

• SeNet_CroReLU achieved a diagnostic accuracy of 98.33% on the experimental dataset.
• The proposed method demonstrated superior performance compared to common neural network models.
• CroReLU showed strong generalization ability on the diverse LC25000 dataset, outperforming existing complex network designs.

Conclusions:

• CroReLU effectively models spatial information and captures critical histological features of lung cancer.
• The activation function significantly improves the accuracy and generalization of deep learning models for lung cancer pathology diagnosis.
• This work represents a novel approach to optimizing deep learning for pathology image analysis via activation function design.