The detection of pulmonary embolism through early diagnosis before severe symptoms appear produces more favorable medical results. The research investigates ViTs as deep learning models to evaluate protein contact maps for their relationship with PE diagnosis. The spatial positions of amino acids are translated into binary matrix maps that originate from protein structural data. The unusual image format does not deter ViTs from using them as visual inputs to discover nepabilities between PE-positive and PE-negative clinical samples. The model utilizing structural biology knowledge and modern deep learning algorithms reached 84.2% accuracy and 82.7% precision in addition to 84.5% recall and 83.6% F1-score making it useful as a computational diagnostic tool.