Accurate localization of anatomical keypoints is fundamental to a variety of clinical and research applications, including acupoint therapy, physical rehabilitation, and biomechanical motion analysis. In this study, we investigate the impact of different deep learning backbone architectures and input image resolutions on the localization accuracy of five clinically important forearm acupoints (LI11, LI10, TE5, LI4, and TE3) using a top-down pose estimation framework. Six backbone models, High-Resolution Network (HRNet), Lite High-Resolution Network (Lite-HRNet), Residual Network (ResNet), Swin Transformer, Visual Geometry Group (VGG) network, and Vision Transformer (ViT), are benchmarked across three input resolutions: 256 × 256, 384 × 384, and 512 × 512 pixels. Localization performance is evaluated using normalized Euclidean distance and its standard deviation, providing a resolution-independent measure of both accuracy and consistency. Among the tested models, HRNet consistently yields the most precise and stable predictions, achieving the lowest average error of 0.009 at the 512 × 512 resolution. Lite-HRNet and ResNet also demonstrate strong performance, particularly at higher resolutions, offering a compelling balance between accuracy and computational efficiency. While ViT exhibits higher error rates at lower resolutions, it significantly improves at 512 × 512, closing the performance gap with CNN-based models. Swin Transformer, however, records the highest average error across all resolutions, indicating its limited suitability for fine-grained localization tasks without additional architectural adaptation. Overall, our findings highlight the importance of backbone selection and resolution scaling in medical keypoint detection tasks. HRNet delivers over 50% lower normalized error compared to the weakest-performing model. These insights provide practical guidance for developing robust, high-precision AI systems for clinical and digital health applications, where anatomical accuracy is essential.