As a first step in treatment, accurate automatic cataract diagnosis is of vital importance. The classification and grading of slit -lamp images can realize the cataract type and severity diagnosis. Deep learning models are widely applied in existing methods. However, a key challenge to improve the performance is the noise existed in medical images. Inspired by the ability of the denoising diffusion probabilistic model to generate noise -robust features in image generation tasks, this work develops a new method for cataract classification to learn and share complementary representations among multiple tasks. To alleviate the existence of general noise, a dual -branch network is proposed to combine the image generation based on the denoising diffusion probabilistic model and the target classification task effectively. A cross fusion module is further designed by two cross attention to enhance the interaction of features generated from two branches. Compared to state-ofthe-art methods, the proposed model improves the performance by a significant margin on three classification datasets and has a more robust tolerance with noise interference. Most notably, for multi -granularity cataract classification, it achieves 73.86% in Recall, 81.18% in Precision, 76.94% in F1 -Score, and 81.79% in Accuracy, which surpasses the performance of the second -place model by 7.43%, 7.07%, 6.76% and 2.37% respectively.