Аннотация:The article discusses the problem of chest X-ray computer tuberculosis diagnosis stability on differ- ent datasets. Some publicly available annotated image datasets are used as examples of such datasets. To eliminate the incompatibility of images from different datasets caused by differences in the acquisition conditions and digital representation, we propose a new dataset of annotated chest X-ray images, namely Sakha-TB dataset, and demonstrate its effectiveness in improving the robustness of neural network meth- ods for tuberculosis diagnosis. The dataset consists of an equal number of single images of healthy and tuberculosis patients obtained using different pieces of equipment. When collecting the data, we focused on creating a sample balanced by gender and diagnoses with close age distributions of patients. Dis- ease diagnosis was carried out by independent double reading with confirmation of positive diagnoses by clinical and laboratory data.We evaluate the performance of some modern image classification algorithms during cross-dataset training and testing on the considered datasets, and present a method for visualizing areas of X-ray images that influenced the results of computer diagnosis. We show that the proposed dataset allows, to some extent, to fill the gaps in the considered datasets and to improve the stability of automatic tuberculosis diagnosis performance for diverse input images.
