This paper presents a deep learning-based Optical Mark Recognition (OMR) system designed to enhance the automation and accuracy of multiple-choice exam correction. Traditional OMR systems rely on rule-based thresholding techniques, which often fail under non-standard conditions such as low-quality scans, variable layouts, or ambiguous marks. In contrast, our approach leverages a fine-tuned EfficientNetB3 convolutional neural network to classify answer bubbles into three categories: confirmed, crossed out, and empty. The model is trained on a real-world dataset using transfer learning and demonstrates superior performance compared to classical OMR tools. A preprocessing pipeline based on OpenCV ensures robust extraction of answer regions. Evaluation results show significant improvements in classification accuracy, highlighting the effectiveness of deep learning in educational assessment automation.