基于深度学习的超声白内障检测算法

Ultrasonic Cataract Detection Algorithm Based on Deep Learning

超声图像中的白内障自动检测是提高诊断效率和性能必不可少的方法。尽管已有使用裂隙灯图像和视网膜图像等其他高质量图像来诊断白内障的方法,但存在设备部署昂贵且基层医疗机构只有超声设备等缺点。此外,处理小规模医学图像仍然具有挑战性。为解决上述问题,提出一种基于检测的深度学习模型。该模型不仅使用YOLOv3来检测焦点区域,而且还同时训练了分类模块,利用DenseNet-161提取的高级特征可以有效地区分白内障的不同程度,从而获得更准确的诊断。由于超声眼球图像没有公开可用数据集,因此我们制作了自己的数据集。我们在收集的数据集上评估了该模型,并达到不低于90%的专家平均检测水平。

Automatic cataract detection in ultrasound images is an essential way to improve the diagnostic efficiency and performance.Although deep learning has worked to diagnose cataract using other high quality pictures like slit-lamp image and retinal images,device deployment is ex⁃pensive and primary medical institutions only have ultrasound equipment.Also,it remains challenging to deal with small-scale medical im⁃age.In this paper,we propose a deep learning model to solve these issues,which not only use YOLOv3 to detect the focal area,but also train a classification module using a DenseNet-161 for high-level feature,which can effectively distinguish different levels of cataract to get more accurate diagnosis.Since there is no publicly available data set for ultrasonic eye images,we made our own data set.We evaluated our model on our collected data set and achieve not lower than 90%,the expert average level of detection.