Pulmonary Embolism Detection with Grad-CAM Visual Attention: A Hybrid Deep Learning-based Approach

v\:* {behavior:url(#default#VML);} o\:* {behavior:url(#default#VML);} w\:* {behavior:url(#default#VML);} .shape {behavior:url(#default#VML);} Normal 0 false false false false EN-GB X-NONE AR-SA /* Style Definitions */ table.MsoNormalTable {mso-style-name:"Table Normal"; mso-tstyle-rowband-size:0; mso-tstyle-colband-size:0; mso-style-noshow:yes; mso-style-priority:99; mso-style-parent:""; mso-padding-alt:0in 5.4pt 0in 5.4pt; mso-para-margin-top:0in; mso-para-margin-right:0in; mso-para-margin-bottom:8.0pt; mso-para-margin-left:0in; line-height:107%; mso-pagination:widow-orphan; font-size:11.0pt; font-family:"Aptos",sans-serif; mso-ascii-font-family:Aptos; mso-ascii-theme-font:minor-latin; mso-hansi-font-family:Aptos; mso-hansi-theme-font:minor-latin; mso-font-kerning:1.0pt; mso-ligatures:standardcontextual; mso-ansi-language:EN-GB;} Background: Pulmonary embolism (PE) is a life-threatening condition and one of the leading causes of mortality, accounting for approximately 60,000100,000 deaths annually in the United States. It occurs when a blood clot blocks pulmonary arteries, disrupting normal blood flow to the lungs. Computed tomography (CT) scans are commonly used for PE diagnosis, typically consisting of 200300 images per patient examination. However, reviewing such a large number of images can be challenging for radiologists, potentially leading to fatigue, reduced focus, and diagnostic errors. Although deep learning (DL) approaches have shown promise in medical image analysis, existing methods often suffer from limited data accessibility, lack of transparency, and issues such as vanishing gradients that affect model generalizability and performance. Methodology: To address these challenges, this study proposes DeepPE, a hybrid deep learning framework for automated PE detection from CT images. The proposed approach integrates DenseNet201 for feature extraction, customized fully connected layers for feature representation, and ResNet50 for weight transformation. In addition, Grad-CAM-based heatmap visualization is incorporated to highlight PE-affected regions and improve interpretability. The model leverages pre-trained networks to enhance performance and reliability while supporting the classification of CT images into nine distinct categories. v\:* {behavior:url(#default#VML);} o\:* {behavior:url(#default#VML);} w\:* {behavior:url(#default#VML);} .shape {behavior:url(#default#VML);} Normal 0 false false false false EN-GB X-NONE AR-SA /* Style Definitions */ table.MsoNormalTable {mso-style-name:"Table Normal"; mso-tstyle-rowband-size:0; mso-tstyle-colband-size:0; mso-style-noshow:yes; mso-style-priority:99; mso-style-parent:""; mso-padding-alt:0in 5.4pt 0in 5.4pt; mso-para-margin-top:0in; mso-para-margin-right:0in; mso-para-margin-bottom:8.0pt; mso-para-margin-left:0in; line-height:107%; mso-pagination:widow-orphan; font-size:11.0pt; font-family:"Aptos",sans-serif; mso-ascii-font-family:Aptos; mso-ascii-theme-font:minor-latin; mso-hansi-font-family:Aptos; mso-hansi-theme-font:minor-latin; mso-font-kerning:1.0pt; mso-ligatures:standardcontextual; mso-ansi-language:EN-GB;} Results: Experimental results demonstrate that the proposed DeepPE framework achieves strong diagnostic performance, including accuracy greater than 96±2%, sensitivity above 91±2%, specificity exceeding 90±2%, and an area under the ROC curve (AUC) above 94±2%. These results indicate that the model effectively identifies pulmonary embolism across multiple CT image classes while providing interpretable visual explanations. Conclusion: The proposed DeepPE model significantly improves the accuracy and reliability of pulmonary embolism detection in CT images. By combining hybrid deep learning architectures with visualization-based interpretability, the framework supports radiologists in identifying PE more efficiently and accurately, thereby enhancing clinical decision-making and reducing the risk of diagnostic errors