CT Diagnostic Mode-Oriented and Cross Difficulty-Aware Network for Pulmonary Embolism Segmentation.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Radiology

Background:

• Accurate segmentation of pulmonary embolism (PE) in computed tomography pulmonary angiography (CTPA) is vital for assessing PE severity and reducing mortality.
• Current segmentation methods struggle with inter-tissue similarities, small PE sizes, and variability in PE shape and location.

Purpose of the Study:

• To develop an advanced network, DMCD-Net, for improved PE segmentation in CTPA.
• To address limitations in existing methods, particularly concerning small PE sizes and tissue similarities.

Main Methods:

• Proposed DMCD-Net utilizes a CT diagnostic mode-oriented approach to differentiate PE from surrounding tissues.
• Incorporated a cross difficulty-aware scheme with cross-supervision and a difficulty-aware loss function to handle challenging PE segmentation cases.
• Evaluated DMCD-Net on two hospital and two public datasets.

Main Results:

• DMCD-Net demonstrated superior performance compared to state-of-the-art methods in PE segmentation.
• The network effectively reduced false positives, especially for small PE and similar tissues.
• Achieved better generalizability across different datasets and clinical settings.

Conclusions:

• DMCD-Net offers a significant advancement in automated PE segmentation for CTPA.
• The proposed methods enhance accuracy and robustness, aiding in precise PE severity assessment.
• This technology has the potential to improve clinical diagnosis and patient outcomes for PE.