Development of an intelligent system for chest X-ray analysis based on convolutional neural networks: segmentation of anatomical structures and automatic calculation of the cardiothoracic index

The article presents the results of developing an intelligent system for automated analysis of chest radiographs, including the stages of semantic segmentation of anatomical structures and calculation of the cardiothoracic index. A comparative analysis of 10 convolutional neural network architectures based on U-Net and SegNet was performed with varying resolution enhancement layers, normalization types, and activation functions, as well as with the application of transfer learning using the pretrained ResNet18 model. Training and evaluation of the models were performed on the open miniJSRT dataset (247 images). The best segmentation results were achieved using the ResUNet model (average IoU = 0.9382). To automatically calculate the cardiothoracic index from the obtained masks, an algorithm was developed that determines the midline and transverse dimensions of the median shadow and the chest. Validation of the system on an independent dataset of 31 high-resolution radiographs showed that the UNetModInstanceNormLeakyReLU model provides the highest accuracy relative to expert measurements (mean absolute error MAE = 0.0169, relative error 4.0 %, Pearson correlation coefficient = 0.898). The obtained data demonstrate the potential of using neural network segmentation to automate cardiothoracic index calculations and serve as the basis for further adaptation of the method to clinical diagnostic tasks.

Authors: P. M. Nurakhunov, A. A. Meldo, I. Yu. Prokhorov

Direction: Informatics, Computer Technologies And Control

Keywords: convolutional neural networks, semantic segmentation, chest X-ray, cardiothoracic index, medical diagnostics, deep learning, U-Net, SegNet, transfer learning, ResNet18

---

View full article