Deepak S, Ameer P M
Department of Electronics & Communication Engineering, National Institute of Technology, Calicut, India.
Department of Electronics & Communication Engineering, National Institute of Technology, Calicut, India.
Comput Biol Med. 2020 Oct;125:103993. doi: 10.1016/j.compbiomed.2020.103993. Epub 2020 Sep 17.
An image retrieval system for medical images aids in disease diagnosis by providing similar images from the medical database to a query image. In this article, a content-based medical image retrieval (CBMIR) system is proposed for the retrieval of magnetic resonance imaging (MRI) images of the brain with three types of tumors:- meningioma, glioma and pituitary tumors. The proposed system uses GoogLeNet encodings via transfer learning as image features. A Siamese Neural Network (SNN), is designed, to represent the GoogLeNet encodings in a two-dimensional (2-D) feature space. The SNN is trained using the contrastive loss function to learn the class-specific image features. The similarity, between a query image and the database images, is measured by the Euclidean metric in the lower dimensional feature space. The proposed method achieves state-of-the-art performance for the retrieval of MRI images with brain tumors. The evaluation is done on the openly available Figshare dataset and the performance metrics used are mean average precision (mAP) and precision@10.
医学图像检索系统通过从医学数据库中为查询图像提供相似图像来辅助疾病诊断。在本文中，提出了一种基于内容的医学图像检索（CBMIR）系统，用于检索患有三种肿瘤（脑膜瘤、胶质瘤和垂体瘤）的脑部磁共振成像（MRI）图像。所提出的系统通过迁移学习使用GoogLeNet编码作为图像特征。设计了一个连体神经网络（SNN），以在二维（2-D）特征空间中表示GoogLeNet编码。使用对比损失函数对SNN进行训练，以学习特定类别的图像特征。在低维特征空间中，通过欧几里得度量来测量查询图像与数据库图像之间的相似度。所提出的方法在患有脑肿瘤的MRI图像检索方面取得了领先的性能。评估是在公开可用的Figshare数据集上进行的，使用的性能指标是平均平均精度（mAP）和精度@10。
