3D Convolutional Neural Network For Ischemic Cardiomyopathy Classification From Single Oxygenation-Sensitive Cardiovascular Magnetic Resonance (OS-CMR) Cine Slice (RF_FR_329)

Ischemic cardiomyopathy (iCMP) is a condition characterized by reduced cardiac function caused by coronary artery disease (CAD), often leading to heart failure. OS-CMR is a promising technique for detecting abnormal variations of myocardial oxygenation and thus indicating CAD (1). Preliminary data indicate a clinical potential for combining deep learning techniques with MRI (2). Convolutional neural network (CNN) architectures may be particularly useful for segmentation and classification tasks (2,3,4). By using a 3D CNN, which is effective in video classification tasks as it considers both spatial and temporal dimensions, it is possible to capture cardiac motion in MRI cine sequences efficiently. We hypothesize that inputting a single mid-ventricular OS-CMR cine slice in a short-axis view into a 3D CNN will result in high discrimination accuracy between iCMP patients and healthy volunteers.

Methods:

 We included patients diagnosed with iCMP and healthy volunteers. CMR imaging was performed using 3T GE, 1.5T GE, and 3T Siemens scanners. For each participant, a single OS-CMR cine slice in a basal-to-mid short axis (SAX) plane with retrospective gating was used. Each cine loop showed 20 phases, which were resized to 255x255 pixels and stacked into a single (20, 255, 255) array. The classification was performed using a 3D CNN (Figure 1). The dataset was split 50/20/30 for training, validation, and testing, with a batch size of 8.

Results:

The dataset consisted of 120 participants, including 60 iCMP patients (53 males and 7 females, mean age 65.0±11.5 years) and 60 healthy volunteers (28 males and 32 females, mean age 54.4±9.0 years). Training and validation loss converged after 20 epochs (Figure 2). The classification accuracy during testing was 0.83, with a sensitivity of 0.94 and a specificity of 0.73 (Figure 3).

Conclusion:

The results demonstrate that even a single slice loop of OS-CMR cine images provides a high classification accuracy in distinguishing between healthy volunteers and iCMP patients. This approach has the potential to automatically flag cine images for the presence of ischemic CMP. Further refinement of the algorithm and studies involving other cardiomyopathies are warranted.