Subgroup-Adaptive Network for Robust Virtual Patient-Specific Quality Assurance

Deep learning-based automatic patient-specific quality assurance (PSQA) alleviates medical resource pressure and ensures the safety of patient treatment plans by predicting the actual dosage difference distribution or metrics. However, deep learning is often affected by extremely unbalanced data distributions, and in clinical practice, the number of normal plans is always much larger than that of abnormal plans. On the other hand, the distribution of patients' lesion sites could be more balanced. The performance of models is often seriously biased in these subgroups, posing a challenge to the reliability of deep learning in the clinical application of PSQA. In this work, we propose a subgroup-adaptive network to address the problem of subgroup drift in PSQA. Our experiments achieved a state-of-the-art (SOTA) MAE of 0.69. Our proposed work can significantly improve the reliability of deep learning in the practical clinical application of PSQA.