Prior Anatomical Knowledge-guided GAN for ICL surgery postoperative prediction based on AS-OCT image

Accurate prediction of postoperative vault, the distance between the implantable collamer lens (ICL) posterior surface and the crystalline lens anterior surface, is critical for the success of ICL surgery. Existing regression-based prediction methods fail to provide visual postoperative observations, which are essential for a comprehensive risk assessment. Anterior segment optical coherence tomography (AS-OCT) enables high-resolution visualization of anterior segment structures. In this work, we pioneer the exploration of using the generative adversarial network to predict postoperative AS-OCT images and quantify the clinical parameter. Given the direct contact between the iris posterior surface and ICL anterior surface, the iris critically influences ICL positioning. Motivated by this, we propose a Prior Anatomical Knowledge-guided GAN (PAK-GAN) to enable both accurate vault prediction and visual observation of postoperative anterior segment structures. Specifically, the Iris morphology perception Auxiliary Branches (IAB) are designed to capture the context of the iris position and shape. Additionally, we incorporate a Gaussian weight map into the loss function to strengthen the model's attention around the iris root region. To validate the performance of our method, we collected paired pre- and post-operative ICL images from multiple hospitals. Comprehensive ablation experiments and comparisons with state-of-the-art methods show that our PAK-GAN achieves the best prediction accuracy for vault and iris position. Furthermore, compared to two clinically used postoperative vault prediction methods, NK-formula and KS-formula, our approach shows the best correlation and agreement with the achieved vault.