Work

Detecting Anatomical Landmarks in 3D Cardiovascular Images Using Convolutional Neural Network

Year: 2020
Type: dissertation
Abstract: Medical imaging enables us to visualize the interior of the body.Traditionally, medical images have been analyzed by doctors, but lately, methods for extracting information automatically from medical images have been explored.Automatical feature extraction is time-saving and makes medical tools more accessible.Convolutional neural networks have proven to be a powerful tool in several medical image processing tasks, with the potential to exceed human performance.In this thesis, we have explored the use of convolutional neural networks for landmark detection in 3D cardiovascular images.Landmark detection in 3D images is useful to provide automatic registration between ultrasound and CT images of the same patient.A patch-based convolutional neural network was used.Two types of network architectures were tested; ResNet18 and a fully convolutional neural network.To improve the performance of the networks, multi-task learning with classification as a secondary task was combined with the landmark detection.In addition, a weighted loss function was applied.The results from the classification were later used to determine the final landmark prediction.The model providing the lowest euclidean error was ResNet18.After post-processing, the average error was 8.78 mm.One application of this method is GE Vingemed's CT-fusion tool.The results were found to provide acceptable accuracy for a semi-automatic landmark detection model.C.1 Results from classification.Regression with regular MSE loss and classification . . . . . . . . . . . . . . .
Author: Betina Høyer Wester
Institution: University of Oslo
Language: English
Open Access status: green
OpenAlex ID: W6981074044
Cites: 21
Cited by: 1
Related to: 9