Malaria remains a pervasive global health challenge, affecting millions of lives daily. Traditional diagnostic methods, involving manual blood smear examination, are time-consuming and prone to errors, especially in large-scale testing. Although promising, automated detection techniques often fail to capture the intricate spatial features of malaria parasites leading to inconsistent performance. In order to close these gaps, this work suggest an improved technique that combines a Self-Attention Mechanism and a Dilated Convolutional Neural Network (D-CNN) to allow the model to effectively and precisely classify malaria parasites as infected or uninfected. Both local and global spatial information are captured by dilated convolutions, and crucial features are given priority by the attention mechanism for accurate detection in complex images. We also examine batch size variation and find that it plays a crucial role in maximizing generalization, accuracy, and resource efficiency. A batch size of 64 produced superior results after testing six different sizes, yielding an AUC of 99.12%, F1-Score of 96, precision of 97.63%, recall of 93.99%, and accuracy of 96.08%. This batch size balances efficient gradient updates and stabilization, reducing overfitting and improving generalization, especially on complex medical datasets. Our approach was benchmarked against existing competitors using the same publicly available malaria dataset, demonstrating a 2-3% improvement in AUC and precision over state-of-the-art models, such as traditional CNNs and machine learning methods. This highlights its superior ability to minimize false positives and negatives, particularly in complex diagnostic cases. These advancements enhance the reliability of large-scale diagnostic systems, improve clinical decision-making, and address key challenges in automated malaria detection.