While automated landmark detection in medical imaging has achieved remarkable accuracy, it still requires sufficiently annotated datasets. This remains a significant barrier to clinical adoption. This paper investigates the effect of the number of annotations available on the model performance without relying on overly specialised few-shot configurations. We explore two practical scenarios: landmark detection with limited annotated data \(({\le }60)\) , and the incremental addition of new landmarks to existing models. Through experiments on hand radiographs, we demonstrate that models trained on just a fraction of the full dataset can achieve an accuracy comparable to that of other methods. Furthermore, we show that new landmarks can be effectively learnt through fine-tuning with as few as five examples, though performance varies with landmark variance. Also, we validate weight initialisation performance and find fine-tuning from prior landmark models tend to under-perform. Our findings suggest that the relationship between the amount of annotated training data and detection accuracy is nonlinear, with diminishing accuracy gains beyond certain thresholds. This insight has important implications for clinical practice, suggesting that label-efficient, modular landmark detection systems are valuable options, particularly when sub-millimetre precision is not critical.