IdFOPNet: Integrating Identity Attention and Fairness Optimization in Anatomical Landmark Detection

Fairness is the key to addressing bias in anatomical landmark point detection. Existing methods tend to ignore individual identity information, resulting in significant bias exhibited between different age groups and genders. Moreover, current studies mainly focus on improving the accuracy of models and lack the dynamic optimisation mechanism of fairness, resulting in the bias problem not being solved effectively. To this end, we propose an anatomical landmark detection network that integrates Identity features and Fairness OPtimization (IdFOPNet). This method leverages the prototype network to detect landmarks by comparing image features with a set of global landmark prototypes. To enhance model fairness, we introduce the Identity Attention mechanism, incorporating identity information as prior knowledge into the detection process. Additionally, we design a penalty-based gradient modulation strategy to dynamically suppress the model’s over-reliance on specific biased information during training. We evaluate the IdFOPNet on the CephAdoAdu and Hand X-Rays datasets. Extensive experimental results demonstrate that our method outperforms SOTA approaches in anatomical landmark detection across different ages and genders, and stays fair as well. Our codes are already released at: https://github.com/ZDX1106/IdFOPNet