Seeking an Optimal Approach for Computer-Aided Pulmonary Embolism Detection.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Pulmonary Medicine

Background:

• Pulmonary embolism (PE) is a life-threatening condition often diagnosed with CT pulmonary angiography (CTPA).
• Deep learning (DL) offers potential for computer-aided diagnosis (CAD) of PE, but numerous competing methods create confusion.
• A systematic comparison of DL approaches for PE diagnosis is needed.

Purpose of the Study:

• To comprehensively analyze and compare competing deep learning methods for computer-aided diagnosis of pulmonary embolism (PE) using CT pulmonary angiography (CTPA).
• To evaluate image-level and exam-level diagnostic performance across various deep learning techniques.

Main Methods:

• Comparison of Convolutional Neural Networks (CNNs) versus Vision Transformers at the image level.
• Evaluation of Self-Supervised Learning (SSL) against supervised learning, and transfer learning versus training from scratch.
• Contrast of Conventional Classification (CC) with Multiple Instance Learning (MIL) at the exam level.

Main Results:

• Transfer learning consistently improved performance for PE diagnosis, outperforming training from scratch.
• Transfer learning combined with SSL demonstrated superior results compared to supervised transfer learning.
• Convolutional Neural Networks (CNNs) generally outperformed Vision Transformers.
• Conventional Classification (CC) showed superior performance over Multiple Instance Learning (MIL) for exam-level diagnosis.

Conclusions:

• Transfer learning, particularly with SSL and CNNs, represents a promising approach for enhancing computer-aided diagnosis of PE.
• The study clarifies the landscape of deep learning methods for PE detection, guiding future research and clinical application.
• Conventional classification methods remain highly effective for exam-level PE diagnosis using CTPA.