Lesion synthesis to improve intracranial hemorrhage detection and classification for CT images

•A novel lesion synthesis method is proposed by applying image inpainting technology.•An automatically Artificial Mask Generator (AMG) is proposed to define artificial masks from a shape pool in unrestricted size, location and shape.•A Lesion Synthesis Network (LSN) is proposed to apply artificial masks to generate hemorrhage lesions on non-lesion images.•A Residual Score Evaluation is presented to select the high reality of the synthetic images. Synthetic images with Residual Scores are combined with real images as the training set.•Experimental results demonstrate the effectiveness of our approach to improve the performances of ICH detection and classification, especially for the micro bleedings. Computer-aided diagnosis (CAD) for intracranial hemorrhage (ICH) is needed due to its high mortality rate and time sensitivity. Training a stable and robust deep learning-based model usually requires enough training examples, which may be impractical in many real-world scenarios. Lesion synthesis offers a possible solution to solve this problem, especially for the issue of the lack of micro bleedings. In this paper, we propose a novel strategy to generate artificial lesions on non-lesion CT images so as to produce additional labeled training examples. Artificial masks in any location, size, or shape can be generated through Artificial Mask Generator (AMG) and then be converted into hemorrhage lesions through Lesion Synthesis Network (LSN). Images with and without artificial lesions are combined for training an ICH detection with a novel Residual Score. We evaluate our method by the auxiliary diagnosis task of ICH. Our experiments demonstrate that the proposed approach can improve the AUC value from 84% to 91% in the ICH detection task and from 89% to 96% in the classification task. Moreover, by adding artificial lesions of small size, the sensitivity of micro bleeding is remarkably improved from 49% to 70%. Besides, the proposed method overcomes the other three synthetic approaches by a large margin.