Unsupervised abnormality detection through mixed structure regularization (MSR) in deep sparse autoencoders.

Area of Science:

• Artificial Intelligence
• Medical Imaging Analysis
• Deep Learning

Background:

• Unsupervised abnormality detection in medical images is crucial for computer-aided detection systems.
• Deep sparse autoencoders offer a promising approach, requiring only healthy data for training.
• Regularization is essential to prevent overfitting in deep learning models.

Purpose of the Study:

• To introduce and evaluate the mixed structure regularization (MSR) approach for deep sparse autoencoders.
• To enhance unsupervised abnormality detection in medical images, specifically coronary artery disease.
• To improve the performance of deep sparse autoencoders by mitigating overfitting.

Main Methods:

• Utilized coronary computed tomography angiography (CCTA) datasets from 90 subjects.
• Trained a deep sparse overcomplete autoencoder model with MSR, incorporating random structure and noise augmentation.
• Compared the performance of the proposed model (SAE-MSR and SDAE-MSR) against standard deep sparse autoencoder (SAE) and deep sparse denoising autoencoder (SDAE) models using tenfold cross-validation.

Main Results:

• The SDAE-MSR model achieved a significant 20% improvement in Area Under the Curve (AUC) and a 30% improvement in Average Precision (AP) for distinguishing mild from severe coronary stenosis.
• The SDAE-MSR demonstrated statistically significant improvements (P < 0.001) in AUC and AP compared to SAE and SDAE.
• Further validation showed an 18% improvement in both AUC and AP for SDAE-MSR in a related task (P < 0.05).

Conclusions:

• Deep sparse autoencoders enhanced with MSR, explicit sparsity regularization, and Gaussian noise corruption show potential for improving unsupervised abnormality detection.
• The proposed MSR approach offers a significant advancement over common deep autoencoder models for medical image analysis.
• This method holds promise for developing more effective deep-learning-based computer-aided detection systems.