基于极坐标拉伸的CT胃癌关键特征增广方法
Enhancing Key Features of Gastric Cancer in CT Modality Using Polar Coordinate Stretching

目的：缓解CT影像的切面空间定义域与神经网络特征提取范围不匹配的问题。资料与方法：在临床中收集了7年跨度内被诊断为胃癌的病人在诊疗过程中生成的895例CT影像资料，清洗出689个高质量序列，由第一名医师确定最大癌症负载横断面并进行标注，并由另一名医师进行审核，组建实验数据集。引入一种基于拉伸的图像增广方法，对齐CT横断面定义域与神经网络特征提取范围，最大化利用CT影像序列的定义域外空间进行数据增广，提高训练效率。结果：在使用所提出的增广方法后，神经元前向响应与反向梯度的均匀度得到提升。同等复杂度下，更少的迭代轮次即能达到更优的效果。结论：所提出的增强方法可以提高神经网络在CT模态上的计算效率，在临床识别任务中有一定应用潜力。
Purpose: To address the mismatch between the slice spatial domain of CT images and the feature extraction range of neural networks. Materials and Methods: A total of 895 CT imaging scans from patients diagnosed with gastric cancer over a span of 7 years were collected in clinical practice. After data cleaning, 689 high-quality sequences were obtained. The maximum cancer load slice was determined and annotated by the first physician and reviewed by another physician to form the experimental dataset. A stretching-based image augmentation method was introduced to align the slice domain of CT with the feature extraction range of neural networks, maximizing the use of the extra-domain space of CT image sequences for augmentation and improving training efficiency. Results: After applying the proposed augmentation method, the uniformity of neuronal forward responses and backward gradients was enhanced. Under the same complexity, fewer iterations were required to achieve better results. Conclusion: The proposed augmentation method can improve the computational efficiency of neural networks on CT modality and has potential for application in clinical cancer detection tasks.