基于可变形卷积的小肠间质瘤多模型联合检测

Multi-model joint detection of gastrointestinal stromaltumor with deformable convolution

小肠间质瘤(GIST)是一种常见的胃肠道肿瘤,目前对GIST的诊断主要依靠CT影像检查。近年来,随着计算机视觉技术在医学影像领域中的广泛应用,极大地提高了诊断效率。针对小肠间质瘤形状大小差异大,病灶区与正常组织器官相似度高的特点,该文提出了一种基于可变形卷积的小肠间质瘤多模型联合检测方案,该方法使用DeepLesion数据集作预训练,在单个检测模型中引入可变形卷积,使网络能够更好地适应小肠间质瘤形态大小差异大的特点,提高特征提取能力。同时对多模型输出的多组检测结果采取非极大值抑制和最高置信度选择的方法选取最高置信度的结果作为模型的输出。在小肠间质瘤数据集上的实验结果表明,该文提出的方法相较于单模型检测在间质瘤检测任务上性能有所提升。

Gastrointestinal stromal tumor(GIST)is a common gastrointestinal tumor.At present,the diagnosis of GIST mainly depends on CT imaging.In recent years,with the wide application of computer vision technology in the field of medical imaging,it can greatly improve the diagnostic efficiency.In view of the large difference in the shape and size of GIST and the high similarity between the lesion area and normal tissues and organs,this paper proposes a multi-model combined detection strategy for gastrointestinal stromal tumor with deformable convolution,which uses the DeepLesion datasets for pre-training,and then,deformable convolution is introduced into a single detection model,so that the network can better adapt to the characteristics of GIST with large differences in shape and size,and improve the ability of feature extraction.At the same time,the method of non-maximum suppression is adopted to select the maximum confidence result from multiple sets of detection results by multiple models as the output of the model.The experimental results on the gastrointestinal stromal tumor data set show that the method proposed in this paper has improved detection performance on stromal tumor detection tasks compared with single-model detection.