University of Twente Student Theses
Deep Learning for Objective Intraoperative Image Analysis during Endovascular Aneurysm Repair : Automatic Artery Detection
Kappe, K.O. (2022) Deep Learning for Objective Intraoperative Image Analysis during Endovascular Aneurysm Repair : Automatic Artery Detection.
PDF
3MB |
Abstract: | Background: Endovascular Aneurysm Repair (EVAR) is the predominant choice for elective and acute treatment of patients with an infrarenal abdominal aortic aneurysm. Completion digital subtraction angiography (cDSA) is performed at the end of an EVAR procedure to evaluate the position and patency of the stent-graft, potential endoleaks and blood flow dynamics. Clinical decision-making is based on the cDSA. However, the evaluation is based on visual inspection by the surgical team and therefore subjective and prone to inter-observer disagreement. In this study, we aim to use deep learning to automatically extract imaging-based features, in particular detection of renal and lumbar arteries, as part of the development of objective intraoperative image analysis. Methods: In chapter 2, we investigated the visibility of lumbar arteries between the pre-operative Computed Tomography Angiography (CTA) and cDSA in thirteen patients treated with EVAR. In chapter 3, a U-Net was trained for the automatic detection of renal arteries on the cDSA, based on two-dimensional (2D) feature projections obtained from the cDSA. The performance of the model was evaluated using the median localization error and successful cumulative detection rate at different distance thresholds and for different combinations of input feature projections. In chapter 4, we developed a deep-learning based method, using U-Net, for the automatic identification of lumbar arteries on the cDSA, using 2D feature projections extracted from the cDSA. The area under the curve of the precision-recall curve (AUCPR) was used to evaluate the performance of the model for different combinations of input feature projections. Results: Over-projection of lumbar arteries by the aortic lumen on the cDSA occurred in 2% of the cases. Invisibility of the lumbar artery on the pre-operative CTA is a good predictor for invisibility on the cDSA, while visibility on the pre-operative CTA was a moderate predictor for visibility on the cDSA. The median localization error of renal arteries for the different input projection combinations on the cDSA were 1.52, 1.52, and 1.65 mm (p=0.853). The successful cumulative detection rate was around 75% for all input projection combinations. The AUCPR was 0.59, 0.61, 0.62 and 0.67 for the identification of lumbar arteries on the cDSA for the different combinations of input feature projections. Conclusion: This study has demonstrated the feasibility of using deep learning models for the automatic detection and identification of renal and lumbar arteries on the cDSA. These extracted features can serve as building blocks for the development of objective intraoperative image analysis during EVAR. |
Item Type: | Essay (Master) |
Faculty: | TNW: Science and Technology |
Subject: | 44 medicine |
Programme: | Technical Medicine MSc (60033) |
Link to this item: | https://purl.utwente.nl/essays/90574 |
Export this item as: | BibTeX EndNote HTML Citation Reference Manager |
Repository Staff Only: item control page