基于自监督学习与语义分割相结合的内窥镜白内障图像分割

Endoscopic cataract image segmentation based on self-supervised learning and semantic segmentation

目的:基于自监督学习与语义分割方法在白内障数据集上训练深度学习模型,分割白内障手术的显微镜图像,以提升算法的准确性和鲁棒性。方法:提出CA-PSP模型实现白内障数据集语义分割,采用自监督模型BYOL预训练模型参数;为了增强网络特征表达能力,在骨干网络加入一个轻量级的网络注意力机制,即坐标注意力模块(coordinate attention)优化学习内容。结果:通过对比实验证明自监督方法及卷积注意力模块对模型性能提升的有效性,像素分割精度为93.9%,Dice系数为76.5%,mIoU系数为64.4%。结论:将自监督学习与语义分割技术相结合并应用在内窥镜白内障图像分割,能有效提升临床诊断的灵活性,为白内障手术阶段的可视化指导提供了有效参考。

Objective Based on self-supervised learning and semantic segmentation methods,deep learning models are trained on the cataract data set,and the microscope images of cataract surgery are segmented to improve the accuracy and robustness of the algorithm.Methods Based on CA-PSP model to achieve semantic segmentation of cataract data sets,adopting self-supervised model BYOL pre-training model parameters;in order to enhance the network feature expression ability,a lightweight network attention mechanism is added to the backbone network,that is,the coordinate attention module(coordinate attention)to optimize the learning content.Results The effectiveness of the self-supervised method and the convolutional attention module to improve the performance of the model is proved through comparative experiments.The pixel segmentation accuracy is 93.9%,the Dice coefficient is 76.5%,and the mIoU coefficient is 64.4%.Conclusion It can effectively improve the flexibility of clinical diagnosis and provide an effective reference for visual guidance in cataract surgical phase through the combination of self-supervised learning and semantic segmentation technology is applied to endoscopic cataract image segmentation.