Bone age assessment is a common clinical method for monitoring children’s growth and development and diagnosing endocrine diseases. However, the traditional Greulich-Pyle (G-P) and Tanner-Whitehouse (TW) methods have several limitations, including reliance on physicians’ experience, complicated procedures, and long processing times. To address this issue, this study proposes a bone age maturity prediction model based on EfficientNet-ATT to improve evaluation accuracy. This model takes EfficientNet-B3 as the backbone network and constructs a bone anatomical attention module (BoneAM) in the basic unit of MBConv, dynamically harmonizing the output of three branches, including channel, spatial, and height-width, through the biological prior and gating mechanism of children’s bone development, and accurately capturing cross-dimensional bone characteristics. Multi-head recurrent spectral attention (MHRSA) is designed within each MBConv block group to alleviate the separation problem of multi-scale bone features and strengthen long-distance dependency modeling. At the same time, the dynamic weighted Huber loss (DWHL) function is adopted to accelerate the convergence with L2 loss when the error is small, and switch to L1 loss to avoid gradient explosion when the error is large, thus improving the robustness of the model to abnormal samples. Experimental results are conducted on the Radiological Society of North America (RSNA) dataset. The results show that the mean absolute error (MAE) and root mean square error (RMSE) of the proposed model reach 4.76 and 6.69 months, respectively, which are 0.73 and 1.01 months lower than those of the benchmark model EfficientNet-B3. Compared with BAA-Net, which currently shows the best performance in bone age assessment models, and ResNet50, which performs best among mainstream algorithms, the MAE is reduced by 0.19 and 0.88 months, and the RMSE is reduced by 0.19 and 0.52 months, respectively. The proposed model achieves superior performance on key error metrics, meeting the requirements of both accuracy and real-time performance in bone age assessment.