Invasive Coronary Angiography remains the standard for diagnosing coronary artery disease, yet the automated classification of stenotic lesions is often hindered by suboptimal image contrast and complex vascular structures. This study evaluates the impact of various image enhancement techniques on the performance of deep convolutional neural network architectures for the binary classification of coronary stenosis. Using the CADICA dataset, a patch-based approach was implemented to distinguish between clinically significant lesions ( \(\ge 50\%\) narrowing) and non-lesion samples. Two phases were conducted: first, a comparative analysis of five benchmark architectures (AlexNet, GoogLeNet, ResNet-18, ResNet-50, and VGG16) as baseline strategy; and second, Gram Equalization, an assessment of four preprocessing algorithms –such as Histogram Equalization, Contrast Limited Adaptive Histogram Equalization, Gamma Correction, and Logarithmic Transformations– on classification accuracy. The results indicate that GoogLeNet achieves the highest baseline performance, achieving an F1-score of 0.705. Furthermore, the findings of Phase 2 demonstrate that while Contrast Limited Adaptive Histogram Equalization significantly benefits deeper architectures such as ResNet and VGG architecture, the baseline GoogLeNet configuration remains the most robust. This research highlights the critical role of preprocessing in clinical decision-support systems and provides a framework to enhance the reliability of deep learning models for stenosis detection in invasive coronary angiography images.