Optimizing Pulmonary Embolism Detection Through Diverse UNET Architectural Variations

A Pulmonary Embolism (PE) is a critical condition that poses a life-threatening risk when blood vessels in the lungs become obstructed. To detect and precisely locate PE, medical professionals typically employ a specialized X-ray technique known as Computed Tomography Pulmonary Angiography (CTPA). Leveraging Deep Learning methodologies, particularly U-shaped encoder-decoder architectures, has emerged as a promising avenue for automating PE segmentation from CTPA images. This research endeavors to assess and compare the performance of several U-shaped networks, including UNET, UNET++, Residual UNET, ARUX, and Attention UNET, in accurately segmenting PE regions. Utilizing a publicly available PE challenge dataset, comprehensive training and testing procedures are conducted, with a meticulous evaluation based on Dice Coefficient, Jaccard Similarity Index, Sensitivity alongside considerations of training time and model parameters. The results of this study put valuable insights into the efficacy and suitability of various deep learning models used for PE segmentation, paving the way for enhanced diagnostic capabilities in clinical settings.