Abstract:
The article describes a solution to the problem of automating the process of segmentation of pulmonary nodules on computed tomography scans to expand the functionality of the previously developed module for determining the size and volume of pulmonary nodules. The main focus of the article is on comparing the accuracy of the models with the ResU-Net, Attention U-Net and Dense U-Net architectures when training on computed tomography images from the LIDC-IDRI dataset in their original form and using two proposed three-channel approaches to their preprocessing. For the three architectures considered, the DSC and IoU values in the ranges 0.8570–0.8735 and 0.7545–0.7881 were achieved. The best metric values were demonstrated by models trained on three-channel images with averaging. In such images, the first channel is represented by a scan in its original form, the second by an averaged scan, and the third by a scan to which anisotropic diffuse filtration is applied. The obtained results allow us to conclude that the use of preprocessing methods is promising for improving the accuracy of segmentation. The article also describes the training of the lung lobes segmentation model using data from the TotalSegmentator dataset. The input data of the modified software module are computed tomography scans, and its output data are processed images and a structured report (DICOM SR). This report, in addition to data on the size and volume of pulmonary nodules, contains information on the lobes in which the detected nodules are located.
Keywords:computer vision, segmentation, lung cancer, lung nodule, computed tomography, medical images, diagnostics, medical decision support system.
Fulltext:
PDF file (926 kB)