The transition module: a method for preventing overfitting in convolutional neural networks

Area of Science:

• Digital pathology and computational imaging.
• Machine learning applications in histopathology.
• Cancer research and diagnostics.

Background:

• Digital pathology enables cost-effective and efficient microscopic disease analysis using digitized slides.
• Accurate identification of complex tumor patterns in digital slides is crucial for cancer research, grading, and burden assessment.
• Convolutional Neural Networks (CNNs) are widely used for analyzing complex histopathological images.

Purpose of the Study:

• To propose a modified deep learning 'transition' module designed to improve generalization in CNNs for digital pathology.
• To enhance the performance of CNNs in classifying complex tumor patterns, particularly when training data is limited.
• To integrate and evaluate the proposed transition module within established CNN architectures like AlexNet and ZFNet.

Main Methods:

• Development of a novel 'transition' module incorporating filters of varying sizes and global average pooling.
• Implementation of the transition module into AlexNet and ZFNet architectures.
• Validation of the modified CNNs on two independent datasets of scanned histology sections for breast tumor classification.

Main Results:

• The inclusion of the modified transition module led to improved performance in breast tumor classification tasks.
• The proposed module demonstrated enhanced generalization capabilities in deep learning frameworks with limited training samples.
• Performance gains were observed when the transition module was integrated into both AlexNet and ZFNet.

Conclusions:

• The novel transition module significantly enhances the performance of CNNs for digital pathology image analysis.
• This approach offers a promising solution for accurate tumor classification in digital pathology, especially in data-scarce scenarios.
• The findings suggest broader applicability of the transition module for improving deep learning models in histopathological image analysis.