基于图搜索和梯度弱化的SDOCT图像ILM层分割

摘要

Accurate detection of RNFL in SDOCT images is of great value for eye disease diagnosis. An automated method of segmenting the upper surface of RNFL is proposed,which is based on 3D graph⁃search algorithm and gradient⁃weakening method. The feature of high contrast between the ILM(internal limiting membrane) surface and the background is taken into ac⁃count and then the gradient⁃weakening method is used to weaken other high gradient areas. With the idea mentioned above,the 3D graph⁃search algorithm can segment the ILM surface accurately. The concept of multi⁃scale segmentation is included during the period of the algorithm,which reduces the running time and memory complexity. Twenty⁃five pairs of eyes’SDOCT images from Stanford University were used to test the method mentioned above. Also the semi⁃automated algorithm of the ITK⁃SNAP soft⁃ware was utilized to make contrast between the experiment results and the software results. The proposed method is fast and feasi⁃ble in 43 cases of the given 50 datasets.%频域光学断层相干扫描（SDOCT）图像针对视网膜中的RNFL层的精确检测，对于一些眼科疾病诊断有很好的参考价值，需要精确分割该层的上表面ILM边界。给出一种3D图搜索与梯度弱化方式相结合的分割ILM表面的方法，提出一种梯度弱化的方法，弱化了图像中其他高对比度区域的梯度，降低其他层对ILM表面分割的干扰，从而在图像中仅有在ILM表面附近较大梯度的前提下，利用经典三维图搜索方法实现对ILM表面的精确分割。分割过程中采用多尺度的分割方法，降低了三维图搜索算法本身时间复杂度和空间复杂度。采用斯坦福大学的25对眼睛的频域光学相干断层（SDOCT）图像，共50组3D图像体数据实验，结合ITK⁃SNAP软件的半自动分割结果进行实验对比。实验结果表明给出的方法对50组数据中的43组数据能够实现ILM表面的快速精确分割。