[1]郑宗生,唐鹏飞,王振华,等.基于改进SOLO_v2的糖尿病黄斑水肿分割模型[J].中国医学物理学杂志,2023,40(1):24-30.[doi:DOI:10.3969/j.issn.1005-202X.2023.01.005]
　ZHENG Zongsheng,TANG Pengfei,WANG Zhenhua,et al.A novel model for diabetic macular edema segmentation based on improved SOLO_v2[J].Chinese Journal of Medical Physics,2023,40(1):24-30.[doi:DOI:10.3969/j.issn.1005-202X.2023.01.005]
点击复制

基于改进SOLO_v2的糖尿病黄斑水肿分割模型()
分享到：

《中国医学物理学杂志》[ISSN:1005-202X/CN:44-1351/R]

卷:
40卷
期数:
2023年第1期
页码:
24-30
栏目:
医学影像物理
出版日期:
2023-01-07

文章信息/Info

Title:
A novel model for diabetic macular edema segmentation based on improved SOLO_v2
文章编号:
1005-202X（2023）01-0024-07
作者:
郑宗生; 唐鹏飞; 王振华; 卢鹏
上海海洋大学信息学院， 上海 201306
Author(s):
ZHENG Zongsheng;  TANG Pengfei;  WANG Zhenhua;  LU Peng
College of Information Technology, Shanghai Ocean University, Shanghai 201306, China
关键词:
糖尿病黄斑水肿; 实例分割; 特征增强; 非极大值抑制
Keywords:
Keywords: diabetic macular edema instance segmentation feature enhancement non-maximum suppression
分类号:
R318；TP391.41
DOI:
DOI:10.3969/j.issn.1005-202X.2023.01.005
文献标志码:
A
摘要:
糖尿病黄斑水肿（DME）是导致糖尿病患者视力损害的常见原因。光学相干断层扫描技术（OCT）有助于增强对糖尿病视网膜病变的早期检测和预防。目前，OCT图像中的DME区域存在大量散斑噪声及小目标区域，现有的实例分割方法存在漏分割等问题。针对上述问题，本文利用特征金字塔转换器（FPT）改进SOLO_v2模型，提出了一种新的DME分割模型（SOLO-OCT），包括：（1）利用基于双域滤波去噪算法去除图像上存在的大量散斑噪声，提高输入图像质量；（2）引入FPT，提高模型对小目标的识别能力和学习能力；（3）改进非极大值抑制（NMS）算法，缓解对小目标区域的漏分割问题。将SOLO-OCT模型与其他实例分割模型（包括Mask R-CNN、SOLO和SOLO_v2）进行了比较，以评估其对DME区域的分割性能。与Mask R-CNN、SOLO和SOLO_v2模型相比，SOLO-OCT模型对DME区域的分割精度（mAP）提高了3.1%，对小目标DME区域的分割精度（APs）提高了2.2%，而单幅图像的处理时间（Fps）只增加了0.009 9 s。本文提出的DME分割模型（SOLO-OCT）可用于大规模糖尿病视网膜病变筛查。
Abstract:
Abstract: Diabetic macular edema (DME) is a common cause of visual impairment in diabetic patients. Optical coherence tomography (OCT) can enhance the early detection and prevention of diabetic retinopathy. At present, there are a lot of speckle noises and small target areas in the DME region in OCT images, and the existing instance segmentation methods have some problems such as missing segmentation. To address the above issues, SOLO_v2 model is improved by feature pyramid transformer, and a novel model (SOLO-OCT model) is proposed for DME segmentation. The proposed method improves the quality of the input image by removing the speckle noises from the image using dual-domain filtering algorithm, and enhances the models ability to recognize and learn small target areas by feature pyramid transformer, and alleviates the problem of missing segmentation for small target areas through improved non-maximum suppression. The SOLO-OCT model is compared with other instance segmentation models (Mask R-CNN, SOLO and SOLO_v2) to evaluate its performance in DME segmentation. Compared with Mask R-CNN, SOLO and SOLO_v2 models, SOLO-OCT model improves the segmentation accuracy of DME region (mAP) by about 3.1% and raises the segmentation accuracy of small-target DME region (APs) by about 2.2%, but the processing time of a single image (Fps) is increased by only about 0.009 9 s. The proposed SOLO-OCT model for DME segmentation can be used for large-scale screening for diabetic retinopathy.

备注/Memo

备注/Memo:
【收稿日期】2022-07-15 【基金项目】国家自然科学基金（41671431）；国家海洋局数字海洋科学技术重点实验室开放基金（B201801034）；上海市科委地方能力建设项目（19050502100）；上海海洋大学科技发展专项基金（A2-2006-20-200211） 【作者简介】郑宗生，博士，副教授，研究方向：深度学习、遥感图像处理，E-mail: zszheng@shou.edu.cn
更新日期/Last Update: 2023-01-07