基于多尺度特征的视盘分割方法

Optic Disc Segmentation Method Based on Multi-scale Features

针对视盘、视杯分割任务中,由青光眼病变引起目标大小显著变化导致的错误分割问题,提出一种使用更轻量级的编码器-解码器网络,并引入金字塔池化模块,通过网络丰富的感受野捕捉更多上下文特征,丰富尺度特征,充分利用全局信息.在数据集RIM-ONE v.3上进行多组对比实验和评估,实验结果表明,该方法对视盘分割的平均交并比为0.908, Dice系数为0.958,均方误差为0.002,比现有算法各项指标性能均有提高.

Aiming at the problem of false segmentation caused by the significant changes in target size due to glaucomatous lesions in the optic disc and optic cup segmentation task, we proposed a lighter encoder-decoder network and introduced pyramid pooling modules to capture more context features, enrich the scale features and make full use of the global information through more receptive fields of the network. A number of comparative experiments and evaluations were carried on the RIM-ONE v.3 dataset, the experimental results show that the mean intersection over union of optic disc segmentation is 0.908, the Dice coefficient is 0.958, and the mean square error is 0.002. Compared with the existing algorithms, the performance of each index is improved.