Advancing Pulmonary Embolism Detection with Integrated Deep Learning Architectures.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Pulmonary embolism (PE) detection from computed tomography (CT) images is critical for patient outcomes.
• Existing deep learning models may require further optimization for accuracy and clinical applicability.
• Hybrid architectures offer potential for enhanced feature extraction in medical image analysis.

Purpose of the Study:

• To introduce HybridNeXt, a novel hybrid deep learning model for biomedical image classification.
• To evaluate HybridNeXt's effectiveness in detecting pulmonary embolism (PE) from CT images.
• To develop a deep feature engineering (DFE) method to further enhance classification performance.

Main Methods:

• A new hybrid convolutional neural network (CNN) architecture, HybridNeXt, was developed, integrating MobileNet, ResNet, ConvNeXt, and Swin Transformer blocks.
• A new dataset for PE detection was created, comprising CT images classified as PE or control.
• A DFE method utilizing multilevel discrete wavelet transform (MDWT), iterative neighborhood component analysis (INCA), and k-nearest neighbors (kNN) was proposed and applied to pretrained HybridNeXt features.

Main Results:

• The HybridNeXt model achieved a test accuracy of 90.14% for PE detection.
• The proposed DFE method, using HybridNeXt features, improved classification accuracy to 96.35%.
• The lightweight version of HybridNeXt demonstrated suitability for clinical applications.

Conclusions:

• The HybridNeXt architecture is highly accurate and effective for biomedical image classification tasks, specifically PE detection.
• The HybridNeXt-based DFE method significantly enhances classification performance.
• The developed models show potential for broader application in various medical image classification challenges.