Lightweight MIDepnet: an Efficient Deep Learning Approach for Myocardial Infarction Detection using Enhanced Image Segmentation

Abstract

Timely identification and detection of Myocardial Infarction (MI) are essential for mitigating cardiac tissue injury and reducing mortality risk. For accurate MI prediction, many algorithms based on Deep Learning (DL) have been created. Amongst, The Encoder-Decoder Convolutional Neural Network (E-D CNN) for image segmentation and a separate CNN model for MI prediction were used in the development of the deep network model for MI detection (MIDepnet). However, the E-D CNN involves many parameters and heavy architectures, hindering real-time image segmentation and detection. To address this, Lightweight E-D CNN (L-E-D CNN) segmentation model is developed for effective MI prediction. In this model, split blocks are introduced in L-E-D CNN which constitutes to two channels. The first channel consists of two parallel paths, one with standard convolution and the other with Multi Dense Dilated Depthwise Separable Convolution (MDDDSC) to capture both local and global features. The use of MDDDSC increase the receptive field with less parameters size. The second channel uses two factorized convolutions with an adaptable dilation rate to increase the receptive field without increasing parameters size. Additionally, a segmented mask training, morphological operations and post-processing pipeline are performed to improve segmentation. The segmented mask training is adopted to learn the inner region with and without papillary muscles and the outer ventricular region to enhance the boundary delineation. The predicted inner mask is subtracted from the outer mask to obtain a refined and separated ventricular region. Morphological operations are then applied to smooth boundaries, fill small gaps and remove minor artifacts. Then, the post-processing stage are applied to improve the boundary clarity, reduces noise and enhances segmentation accuracy. Lastly, the CNN-LSTM model is used for MI prediction using the features obtained from the L-E-D CNN. The complete model is termed as Lightweight MIDepnet (LMIDepnet). Finally, the outcomes show that the LMIDepnet algorithm on the Hamad Medical Corporation, Tampere University and Qatar University (HMC-QU) dataset and Realistic Synthetic 2D Ultrasound dataset obtains the accuracy of 95.97% and 94.89%, respectively, compared to the UNet + Support Vector Machine (SVM), UNet + CNN, DeepLabV3 + ResNet50, UNet + DenseNet121, SegNet + Saimese Neural Network (SNN) models.