LayNet-A multi-layer architecture to handle imbalance in medical imaging data

Area of Science:

• Machine Learning
• Deep Learning
• Medical Informatics

Background:

• Imbalanced datasets pose challenges for traditional machine learning classifiers, leading to minority class underestimation.
• Inaccurate minority class prediction in medicine can result in critical false negatives for diseases like Covid-19.

Purpose of the Study:

• To introduce LayNet, a multilayer deep learning architecture designed to mitigate class imbalance issues.
• To improve the classification accuracy of minority classes in highly imbalanced medical datasets.

Main Methods:

• LayNet divides classes across multiple layers, creating a hybrid class for minor classes in higher layers.
• Each layer employs a distinct model to classify inputs as either a singleton or hybrid class.
• A method for distributing classes across architectural levels is presented.

Main Results:

• The proposed LayNet architecture demonstrated superior performance compared to traditional single-layer models.
• Evaluation on Ocular Disease Intelligent Recognition, Covid-19 Radiography, and Retinal OCT datasets confirmed LayNet's effectiveness.
• LayNet achieved a more balanced class distribution at each layer, enhancing minority class recognition.

Conclusions:

• LayNet offers a promising solution for deep learning classification in imbalanced medical datasets.
• The multilayer approach effectively addresses the limitations of single-layer networks in handling class imbalance.
• Improved minority class detection via LayNet can lead to more accurate medical diagnoses and better patient outcomes.