医学图像语义概念识别方法研究

医学图像语义概念识别是医学图像知识表示的重要技术环节。研究医学图像语义概念识别方法,有助于机器理解和学习医学图像中的潜在医学知识,在影像辅助诊断和智能读片等应用中发挥重要作用。将医学图像的高频概念识别问题转化为多标签分类任务,利用基于卷积神经网络的深度迁移学习方法,识别有限数量的高频医学概念;同时利用基于图像检索的主题建模方法,从给定医学图像的相似图像中提取语义相关概念。国际跨语言图像检索论坛ImageCLEF于2018年5月组织ImageCLEFcaption 2018评测,其子任务“概念检测”的目标是给定222 314张训练图片和9 938张测试图片,识别111 156个语义概念。上述两种方法的实验结果已被提交。实验结果表明,利用基于卷积神经网络的深度迁移学习方法识别医学图像高频概念,F1值为0.092 8,在提交团队中排名第二;基于图像检索的主题模型可召回部分低频相关概念,F1值为0.090 7,然而其性能依赖于图像检索结果的质量。基于卷积神经网络的深度迁移学习方法识别医学图像高频概念的鲁棒性优于基于图像检索方法的鲁棒性,但在大规模开放语义概念的识别技术研究上仍需进一步完善。

Research on Methods of Semantic Concept Detection from Medical Images

Identifying useful concepts from large scale medical images is an important technology for image knowledge representation. Developing semantic concept detection algorithms is helpful to machine understanding and learning latent knowledge from medical images, and plays an important role in image-assisted diagnosis and intelligent image reading. In this study, the problem of detecting high-frequency concepts from medical images was transformed into a multi-label classification task. The deep transfer learning method based on convolutional neural network (CNNs) was used to recognize high-frequency medical concepts. The image retrieval-based topic modeling method was used to obtain the semantically related concepts from the similar images of given medical images. The CLEF cross language image retrieval track (ImageCLEF) launched the ImageCLEFcaption 2018 evaluation task on May, 2018, in which the Concept Detection subtask identified 111,156 semantic concepts from 222,314 training images and 9,938 test images. Our research group presented experimental results of both methods. The CNNs-based deep transfer learning method achieved the F1 score of 0.0928, which ranked second in all the submission teams. The retrieval-based topic model could recall some low-frequency concepts and achieved the F1 score of 0.0907, but dependent heavily on the image retrieval results. The results proved that the CNNs-based deep transfer learning method showed preferable robustness on high-frequency concept detection, but still a lot of room for improvement in the research of large-scale open semantic concept detection.