Automated Detection of Pulmonary Embolism from Computed Tomography Scans Using Deep learning
Code: G-1606
Authors: Parastoo Honarvar Namin * ℗, حسن قبادی مراللو, رنا جنتی, اسما سلمانی, محمد علی جواد زاده
Schedule: Not Scheduled!
Tag: Biomedical Signal Processing
Download: Download Poster
Abstract:
Abstract
Background and aims: Since pulmonary embolism is a dangerous and life-threatening disease and the use of lung computed tomography presents challenges, employing modern methods and computer-aided diagnosis can accelerate the treatment and diagnostic process while reducing the workload of radiologists. The primary goal of this study is to utilize deep learning techniques to enhance computer-assisted detection of pulmonary embolism. To achieve this, we propose a classification-guided detection approach that effectively leverages probabilistic inference from the classifier to guide detection predictions, contributing a novel advancement in the field of automated embolism diagnosis. Method: Our classification system incorporates an attention-guided convolutional neural network (AG-CNN) that utilizes local texture information through an attention mechanism. This approach mimics the human expert method, which first considers the overall appearance of the image and then focuses on local regions before making a decision. Results: The classifier demonstrates strong performance on our dataset, achieving an area under the receiver operating characteristic curve (AUROC) of 0.927, a sensitivity of 0.862, a specificity of 0.879, and an F1-score of 0.805 with the Inception v3 backbone architecture. Furthermore, AG-CNN outperforms the DenseNet-121 baseline model, achieving a notable 1.8% improvement in AUROC. While previous research has primarily focused on detecting embolisms in major arteries, our use of advanced object detection models and hybrid techniques significantly improves the accuracy of detecting small embolisms in peripheral arteries. Finally, our proposed classification-guided detection approach further refines diagnostic metrics and contributes to new advancements in the field, achieving mAP50, sensitivity, and F1-score values of 0.846, 0.901, and 0.779, respectively outperforming previous benchmarks. With a substantial 3.7% improvement in mAP50. Conclusion: our research aims to enhance patient care for pulmonary embolism by integrating AI solutions into clinical workflows, highlighting the potential for human-AI collaboration in medical diagnosis.
Keywords
,Computed Tomography,Deep learning, Pulmonary Embolism
