Pulmonary embolism (PE) represents a perilous health crisis where delays in diagnosis can severely impact patient outcomes. Nevertheless, the quest for automated PE detection through CT pulmonary angiography (CTPA) faces obstacles such as anatomical variability, imaging artefacts, and imbalances within datasets. This research unveils an innovative framework designed to tackle these hurdles, featuring a sophisticated preprocessing pipeline alongside a hybrid deep learning architecture. The established pipeline integrates deep learning facilitated pulmonary localization (94.2% IoU) alongside image registration methodologies to standardize CTPA imaging, thereby minimizing variability and enhancing the reliability of feature consistency. This preprocessing phase has proven to be vital, establishing a foundation for a hybrid EfficientNet-V2-S Transformer model. This fuses the power of convolutional networks for spatial feature extraction with transformer-based attention to capture global context. Tested on 19,309 CTPA images from the RSNA Challenge dataset, our comprehensive system attained an accuracy of 99.21%, sensitivity of 99.45%, and specificity of 98.97%. The principal contribution of this endeavour lies in showcasing how a robust preprocessing pipeline can significantly enhance performance, elevating baseline accuracy from 89% to cutting-edge results. This study offers an effective and adaptable strategy for automated PE detection, holding substantial promise to enhance clinical workflows within emergency radiology.