The accurate selection of healthy sperm is a major component in the success of assisted reproductive technology (ART). However, traditional approaches often struggle to properly assess DNA integrity, a vital parameter for determining a sperm cell’s reproductive potential. This challenge is compounded by laboratory inconsistencies and the difficulties in standardizing testing procedures. Additionally, traditional semen analysis methods are both time-consuming and costly, underscoring the need for efficient alternatives. In this study, we present a computationally efficient deep learning approach designed to identify sperm cells with high DNA integrity. We employ a lightweight convolutional neural network with a regression head, tailored specifically for this task. To optimize the model for deployment on edge devices, an 8-bit quantization process is applied, significantly reducing the model’s size and computational demands. The proposed model is rigorously evaluated on a public dataset, achieving Pearson correlation scores of 0.625 in donor-independent experiments. Both the original CNN model and its quantized counterpart demonstrate state-of-the-art performance on the dataset, highlighting their potential for real-world applications. We believe this novel approach to DNA fragmentation analysis will greatly enhance the efficiency and accuracy of sperm selection in microscopy-based ART, offering a promising solution for broader clinical adoption.