The structural evaluation of cardiac Magnetic Resonance Imaging (MRI) region is essential in diagnosing cardiomyopathy, myocardial infarction, and ventricular abnormalities. Evaluating features in the right ventricular (RV), left ventricular (LV), and myocardium (MYO) regions are used for the diagnosis. Conventional U-Net systems struggle to identify regions with similar shape variations accurately. To address this, the study proposes an improved deep learning architecture to segment multiple regions at the end-diastole (ED) and end-systole (ES) MRI phases. This research has also coupled the modified 2D U-Net design with structural modifications in the decoding layer, incorporating deep supervision loss. Further, the model leverages an attention gate to prioritize feature maps based on the relevance of regions. A frame difference method is used to recover the Region of Interest (ROI) from the MRI frames before segmentation to reduce the training complexity and the incorrect predictions. The study also uses various loss functions to explore the impact of class imbalance between the target and background regions. Hybrid loss and Focal Tversky Loss (FTL) are proposed to enhance region segmentation and an ablation study is conducted to compare the performance of different loss functions. Experimental results using the Automated Cardiac Diagnosis Challenge dataset show dice coefficients of 0.968 (LV), 0.893 (MYO), and 0.935 (RV) at ED and 0.932 (LV), 0.90 (MYO), and 0.91 (RV) at ES phases. The model also produced an IoU score of 0.963 (LV), 0.791 (MYO), and 0.938 (RV) at ED and 0.875 (LV), 0.816 (MYO), and 0.792 (RV) at ES phases. These results show that the proposed model outperformed U-Net and other segmentation methods by significantly improving the segmentation results, specifically on the myocardial and right ventricle regions at the ES phase.