| Tartalmi leírás: |
Cardiovascular diseases (CVDs) are among the leading causes of deaths worldwide. Most of these diseases are difficult to diagnose in time; however, many of the lesions connected to them can be seen on PA chest radiographs. Although this is not the primary modality for the identification of such lesions, it is the most commonly used in daily medical practice. In a previous research we proposed a method for the accurate and robust segmentation of the heart [1]. This method uses a classical image processing algorithm to localize the lung area as a preprocessing step. In this paper, we present a deep neural approach to tackle this problem aiming for a more accurate segmentation. This is necessary, as a more accurate lung localization model may improve the performance of our heart segmentation model, furthermore if both accurate and robust heart and lung segmentations are available, it is possible to calculate the cardiothoraic ratio (CTR), which is a common metric used to decide whether a patient has cardiomegaly. In our research we train and assess a modified UNet architecture on the JSRT, Montgomery and Shenzen datasets and qualitatively evaluate the model on a private dataset coming from daily medical practice. The last step is important, as publicly available datasets often lack the variance and come from a different distribution than the data acquired in daily medical in practice.
|
