Improving Myocardial Infarction Detection from Echo Images using Contrastive Guided Adversarial Denoising Diffusion Probabilistic Model
Belavendhiran Arockia Valanrani, S. Suganya · International Journal of Engineering Trends and Technology · 2024
Early detection and diagnosis of Myocardial Infarction (MI) are crucial for preventing cardiac damage or death. Deep Learning (DL) methods effectively diagnose MI, but data scarcity is a primary challenge. Generative Adversarial Networks (GAN) models provide sufficient images by generating quality echo images. However, the diversity of the images generated by GAN is limited due to its predominant usage in generating and translating images between different sources. Low diversity leads to degrade the performance of deep learners in MI diagnosis. To solve this, the Contrastive Guided Adversarial Denoising Diffusion Probabilistic Model with GAN (CGADDPM-GAN) model is proposed in this paper to generate high-quality echocardiography images with high diversity for efficient MI detection. The combination of CGADDPM and GAN assisted in learning the Reverse Denoising Task (RDT) to represent the important anatomical features in produced image samples. In CGADDPM, the Diffusion Probabilistic Model (DPM) is used to generate samples that coincide with data within a limited range for network parameterization. Contrastive Learning Loss (CLL) is integrated with DPM to improve the quality of learned representations through a learnable nonlinear transformation. Representation Learning (RL) is introduced with Contrastive Learning (CL) to enhance the performance through normalized embeddings and parameter adjustments, resulting in smaller batch sizes. The synthesized images from CGADDPM-GAN are fed into an Encoder-Decoder Convolutional Neural Network (E-D CNN) for segmentation. The features from segmented images are fine-tuned through feature engineering. The fine-tuned features are then utilized in CNN for training and predicting MI. The complete framework is named the Deep network model for MI detection (MIDepnet), which provides synthesized echocardiography images with a large diversity and high accuracy in MI detection.