A Novel Computer-Aided Diagnosis Scheme on Small Annotated Set: G2C-CAD.

Area of Science:

• Medical Imaging Analysis
• Machine Learning in Healthcare
• Computer-Aided Diagnosis (CAD)

Background:

• Computer-aided diagnosis (CAD) systems are crucial for improving diagnostic accuracy in medical imaging.
• A significant challenge in developing effective CAD systems is the scarcity of sufficient labeled training samples.
• Generative Adversarial Networks (GANs) offer a potential solution for data augmentation in machine learning.

Purpose of the Study:

• To address the limitation of insufficient labeled samples in medical image analysis for CAD.
• To develop and evaluate a novel semisupervised learning algorithm, G2C-CAD, leveraging GANs.
• To improve the classification performance of pulmonary nodule signs related to lung cancer.

Main Methods:

• Utilized a Deep Convolutional GAN (DCGAN) framework to generate simulated unlabeled samples from a small labeled dataset.
• Employed a CNN-based fuzzy Co-forest classifier trained iteratively with generated unlabeled data for semisupervised learning.
• Evaluated the G2C-CAD system against C4.5 random decision trees and a non-fuzzy G2C-CAD approach using ROC analysis and confusion matrices.

Main Results:

• The G2C-CAD system achieved high AUCs for classifying lung cancer signs, including noncentral calcification (0.946), lobulation (0.912), spiculation (0.908), and nonsolid/ground-glass opacity (GGO) texture (0.887).
• G2C-CAD demonstrated an average accuracy improvement of 14% over the C4.5 random decision tree.
• On the LISS dataset, G2C-CAD achieved excellent AUCs for GGO (0.972), lobulation (0.964), spiculation (0.941), and pleural indentation (0.967).

Conclusions:

• G2C-CAD effectively addresses the challenge of limited labeled data in medical image analysis for CAD.
• The developed system can facilitate the creation of comprehensive training sample libraries for CAD classification.
• G2C-CAD represents a promising approach for advancing future medical image analysis and diagnostic tools.