The localization of landmarks plays a critical role in medical image analysis, particularly in tasks like cephalometry which involves the radiographic examination of the human cranium. This paper proposes a method for identifying anatomical points in cephalometric images using pre-trained models to address the challenge of insufficient data, a major obstacle in medical image training. A patch classification model is developed utilizing deep feature maps extracted from renowned pre-trained models such as VGGNet16, ResNet50, InceptionNetV3, and DenseNet121, trained on the ImageNet dataset with the exclusion of the last dense layers. These feature maps are then classified using the Light Gradient Boosting Machine algorithm. Landmark coordinates are extracted from patches that get the highest probability value in patch classification for containing the desired anatomical point. Experimental evaluation conducted on a dataset comprising 300 cephalometric images for 19 anatomical landmarks demonstrates average identification accuracies of 81.78%, 78.43%, 68.11%, and 77.50% for VGG16, ResNet50, InceptionNetV3, and DenseNet121, respectively. Additionally, the gradient-based explainability method is employed on the most effective VGGnet16 feature-extracting model to discern the image attributes extracted from patches, using ImageNet training weights, which are particularly conducive to prioritizing landmark localization.