This research endeavors to automate the detection of pulmonary embolism in lung perfusion scintigraphy images using image processing and artificial intelligence methods, aiming to assess embolism levels through localization and various statistical calculations. The study utilizes CT and perfusion images from 37 individuals, with 20 diagnosed with pulmonary embolism and 17 classified as normal. Employing a threshold of −150 HU (Hounsfield Unit) and morphological operations, including opening and closing with a 5-voxel disc-shaped structuring element, the lungs are segmented to remove air and blood veins. Alignment of CT and perfusion images is achieved through affine transformation and bilinear interpolation. The segmented lungs serve as masks on perfusion images, from which intensity-based features are extracted comprising a total of 23 features used in the analysis. In the Multi-Layer Perceptron (MLP) model, training phase metrics indicate mean sensitivity, specificity, accuracy, and F1 Score at 94.41%, 97.44%, 95.94%, and 95.64%, respectively. Testing phase results for MLP exhibit comparable metrics at 94.44%, 95.24%, 94.87%, and 94.44%. Conversely, the Support Vector Machine (SVM) model demonstrates distinctive performance characteristics, with mean sensitivity during training at 98.41%, decreasing to 64.44% in testing. Mean specificity decreases from 95.06% in training to 84.92% in testing, and mean accuracy declines from 95.94% to 75.52%. Mean F1 Score follows a similar trend, dropping from 95.64% in training to 70.09% in testing, emphasizing the diverse performance outcomes of the two models across training and testing datasets.
Veritabanları (dc.source.platform) | Scopus |
Department (dc.contributor.department) | Elektrik/Elektronik Mühendisliği |
Yazar (dc.contributor.author) | Gökalp Tulum |
Yazar (dc.contributor.author) | Onur Osman |
Tür (dc.type) | Konferans Bildirisi |
Eser Adı (dc.title) | Automated Detection of Pulmonary Embolism Using CT and Perfusion Spectral Images |
Konu Başlıkları (dc.subject) | Artificial intelligence |
Konu Başlıkları (dc.subject) | Image processing |
Konu Başlıkları (dc.subject) | Lung Perfusion Scintigraphy |
Konu Başlıkları (dc.subject) | Pulmonary Embolism Detection |
Yayın Yılı (dc.date.issued) | 2024 |
Yayıncı (dc.publisher) | Springer Science and Business Media Deutschland GmbH |
Kitap Adı (dc.identifier.kitap) | Lecture Notes in Networks and Systems |
ISSN (dc.identifier.issn) | 23673370 |
Açık Erişim Tarihi (dc.date.available) | 2040-01-01 |
Tek Biçim Adres (dc.identifier.uri) | https://hdl.handle.net/20.500.14081/2151 |
Özet (dc.description.abstract) | This research endeavors to automate the detection of pulmonary embolism in lung perfusion scintigraphy images using image processing and artificial intelligence methods, aiming to assess embolism levels through localization and various statistical calculations. The study utilizes CT and perfusion images from 37 individuals, with 20 diagnosed with pulmonary embolism and 17 classified as normal. Employing a threshold of −150 HU (Hounsfield Unit) and morphological operations, including opening and closing with a 5-voxel disc-shaped structuring element, the lungs are segmented to remove air and blood veins. Alignment of CT and perfusion images is achieved through affine transformation and bilinear interpolation. The segmented lungs serve as masks on perfusion images, from which intensity-based features are extracted comprising a total of 23 features used in the analysis. In the Multi-Layer Perceptron (MLP) model, training phase metrics indicate mean sensitivity, specificity, accuracy, and F1 Score at 94.41%, 97.44%, 95.94%, and 95.64%, respectively. Testing phase results for MLP exhibit comparable metrics at 94.44%, 95.24%, 94.87%, and 94.44%. Conversely, the Support Vector Machine (SVM) model demonstrates distinctive performance characteristics, with mean sensitivity during training at 98.41%, decreasing to 64.44% in testing. Mean specificity decreases from 95.06% in training to 84.92% in testing, and mean accuracy declines from 95.94% to 75.52%. Mean F1 Score follows a similar trend, dropping from 95.64% in training to 70.09% in testing, emphasizing the diverse performance outcomes of the two models across training and testing datasets. |
Orcid (dc.identifier.orcid) | 0000-0003-1906-0401 |
Orcid (dc.identifier.orcid) | 0000-0001-7675-7999 |
Dil (dc.language.iso) | En |
ISBN (dc.identifier.isbn) | 9783031628702 |
DOI (dc.identifier.doi) | 10.1007/978-3-031-62871-9_28 |
Araştırma Alanı (dc.relation.arastirmaalani) | Engineering |