Segment-based and Patient-based Segmentation of CTPA Image in Pulmonary Embolism using CBAM ResU-Net

An accurate pulmonary embolism segmentation from computed tomography pulmonary angiography (CTPA) images is very important in pulmonary embolism diagnosis. However, the segmentation process nowadays is done manually by physicians, making the process time-consuming and the accuracy dependent on the physician's ability. Hence, we focused on improving the accuracy of an automatic segmentation method by employing image preprocessing techniques and modifying U-Net architecture in many ways. This paper introduced ResU-Net (residual U-Net) and CBAM ResU-Net (ResU-Net with the addition of CBAM attention modules in the skip connections) models to this framework. Furthermore, we explored two training approaches: segment-based and patient-based. Our experimental results showed that segment-based models performed substantially better than patient-based models. Regardless of training approaches, the most accurate model was CBAM ResU-Net, which achieved up to 0.8719 in Dice score for the segment-based approach and 0.6751 for the patient-based approach.