关键词: 人工智能(AI), 磁共振成像(MRI), 样本增广, 批量标注, 影像辅助诊断

Abstract: Training data enrichment is a key factor in artificial intelligence (AI) technology development. At present, the bottleneck problem is that the quantity and type of labeled training data in valid samples are unable to meet the requirements of AI+MRI aided diagnosis. In this paper, an effective approach to solve the problem was presented. High resolution isotropic multi-dimensional data of regions of interests from patients or healthy volunteers were first acquired via a series of scanning on clinical MRI scanners, including quantitative T₁, T₂, proton density (Pd) and apparent diffusion coefficient (ADC) measurements. These data were then used as the ground truth, from which different types of images associated with different imaging sequences and parameters were obtained with a virtual MRI technology. The type of the images with the best boundary resolution were then selected manually by experienced doctors, on which three-dimensional mask matrix was obtained by manual contouring and labeling, serving as the template for other types of images. This enrichment method was developed as a software platform, which could provide sufficient quantity of image data from a small number of positive cases, thus meeting the data training enrichment requirement of AI+MRI diagnosis at low cost and with high efficiency.

Key words: artificial intelligence (AI), magnetic resonance imaging (MRI), training data enrichment, batch labeling, image aided diagnosis