Conditions like angina, heart attack, heart failure, and arrhythmias can be life-threatening, often linked to or caused by stenosis. Manual diagnosis is time-consuming, but automation can improve speed and accuracy. Hence, this work presents a custom Attention U-Net architecture for stenosis detection from X-ray angiography images. The U-Net is widely used for medical image segmentation but struggles with small or thin structures like blood vessels. To achieve clear segmentation results and precise boundaries, an attention mechanism is employed in this work to minimize the impact of background information and highlight important features. Although the attention block increases the model performance, it also increases the load on the system, which in turn increases the model training time. To address this, the proposed work introduces residual block to reduce complexity and enhance feature propagation further stabilizing deeper training and allowing the network to benefit from fine-grained focus and strong gradient flow. The model pipeline incorporates deep supervision in the middle and interpretation layers, directly training them with auxiliary loss functions, which makes the model more robust. These coordinated components result in improved training efficiency and performance as compared to the isolation in the state-of-the-art. The model is evaluated using F1-score, accuracy, and IoU, achieving results of 93.2%, 99.83%, and 88.7%, respectively. The proposed approach shows significant improvements in F1-score and IoU (47.2% and 89.4%, respectively) compared to the baseline U-Net model, with accuracy remains comparable.