Weakly supervised attention model for RV strain classification from volumetric CTPA scans

Area of Science:

• Cardiology
• Radiology
• Artificial Intelligence in Medicine

Background:

• Right ventricle (RV) evaluation is crucial for cardiovascular and pulmonary disorders.
• Pulmonary embolism (PE) diagnosis and risk stratification are critical due to high mortality rates.
• High-risk PE is linked to RV dysfunction from acute pressure overload, necessitating accurate classification.

Purpose of the Study:

• To develop an automated method for RV strain classification from computed tomography pulmonary angiography (CTPA) scans in PE patients.
• To improve early diagnosis and risk stratification for high-risk PE.
• To leverage CTPA scans for both PE diagnosis and RV strain assessment.

Main Methods:

• Utilized a 3D DenseNet network architecture enhanced with residual attention blocks.
• Employed weak labels extracted from CTPA scan reports for RV strain classification.
• Developed a fully automated, end-to-end trainable network requiring minimal preprocessing and labeling.

Main Results:

• Achieved an Area Under the Receiver Operating Characteristic Curve (AUC) of 0.88 for RV strain classification.
• Demonstrated a sensitivity of 87% and specificity of 83.7% using Youden's index.
• Outperformed existing state-of-the-art 3D Convolutional Neural Networks (CNNs) in RV strain classification.

Conclusions:

• A small dataset of unmarked CTPAs can effectively be used for RV strain classification.
• This is the first study to address RV strain classification from CTPA scans using this specific deep learning architecture.
• The proposed self-attention blocks offer a generalizable methodology for 3D medical image classification tasks.