Abstarct: Long-Leg Radiographs (LLR) are considered the gold standard for assessing lower limb alignment and play a crucial role in planning orthopedic treatments. Accurate evaluation of lower limb abnormalities requires the precise localization of anatomical landmarks, which are essential for analyzing physiological axes and measuring alignment indices. . In addition to being time-consuming, existing manual methods are prone to human errors and limited reproducibility. Key challenges in deep learning models include limited access to medical datasets, patient morphological diversity, and the presence of orthopedic joint implants. Accurate localization of anatomical landmarks in LLR images is critical for automating this process. Current single-stage methods often fail to provide sufficient accuracy in landmark localization, while multi-stage models relying on bone segmentation are hindered by boundary errors and difficulties in accurately identifying bone lines. In this research, a two-stage deep learning-baxsed model is proposed for automatic localization of anatomical landmarks and lower limb alignment assessment in LLR images. The system is designed by enhancing the YARLA model and utilizing the YOLOv8n algorithm for joint detection, with ResNet-50 employed for precise anatomical landmark localization of five alignment indices (HKA, LDTA, JCLA, mMPTA and mLDFA). One of the key innovations of this research is the design of a custom loss function that eliminates the need for post-processing by modifying the dimensions of YOLOv8n bounding boxes, reducing computational overhead. Additionally, combining the Mean Absolute Error (MAE) and SMOOTHL1 as the loss function significantly improves landmark localization accuracy, while data augmentation techniques enhance model performance, addressing the challenges posed by data scarcity and patient morphological diversity. For performance evaluation, a Bland-Altman comparative analysis with radiologists’ observations was conducted. The results indicated that the deviation from reference values for all alignment indices was less than 2 mm, with a measurement time of approximately 4 seconds. The mean absolute error of 0.2 degrees for the HKA index in the proposed model was significantly lower than values reported in previous studies, proving the high efficiency of this approach. Trained on a dataset of 422 LLR images, which included 26 anatomical landmarks with the right leg convention, the model achieved high accuracy in analyzing patients with various conditions such as osteotomy, arthroplasty, and knee osteoarthritis, offering up to 15% higher accuracy compared to existing methods. Despite challenges such as low image contrast and limited training data, this study provides an innovative deep learning architecture that offers an effective solution for reducing human error, improving clinical evaluation quality, and facilitating orthopedic surgery planning.