Convolutional Neural Networks for the Detection of Diseased Hearts Using CT Images and Left Atrium Patches

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Cardiology

Background:

• Cardiovascular disease is a major cause of mortality in the US.
• Accurate identification of cardiac disease on 3D CT scans is clinically significant.
• Automated methods can enhance diagnostic efficiency and accuracy for 3D CT imaging.

Purpose of the Study:

• To develop and implement a convolutional neural network (CNN) model for identifying diseased hearts on conventional 3D CT images.
• To compare the performance of 2D and 3D CNN models in classifying cardiac health.
• To evaluate the diagnostic capability of the proposed CNN method.

Main Methods:

• Chest CT scans were acquired for six healthy patients and six patients with prior cardiovascular disease.
• Left atria were segmented, and 2D and 3D image patches were generated.
• Separate CNNs were trained using leave-one-out cross-validation, with 3D patches yielding superior results.

Main Results:

• The 3D CNN model achieved a higher testing accuracy compared to the 2D model.
• The final average area under the curve (AUC) was 0.840 ± 0.065.
• The average classification accuracy for distinguishing healthy from diseased hearts was 78.9% ± 5.9%.

Conclusions:

• CNN-based analysis of 3D CT images can effectively differentiate healthy hearts from those with cardiovascular disease.
• The developed automated method shows promise for improving the diagnosis of cardiac conditions.
• This approach offers a valuable tool for clinical applications using conventional 3D CT data.