Sie befinden Sich nicht im Netzwerk der Universität Paderborn. Der Zugriff auf elektronische Ressourcen ist gegebenenfalls nur via VPN oder Shibboleth (DFN-AAI) möglich. mehr Informationen...
Multi-task Learning for Hierarchically-Structured Images: Study on Echocardiogram View Classification
Ist Teil von
Simplifying Medical Ultrasound, p.185-194
Ort / Verlag
Cham: Springer Nature Switzerland
Quelle
Alma/SFX Local Collection
Beschreibungen/Notizen
Echocardiography is a crucial and widely adopted imaging modality for diagnosing and monitoring cardiovascular diseases. Deep learning has been proven effective in analyzing medical images but is limited in echocardiograms due to the complexity of image acquisition and interpretation. One crucial initial step to address this is automatically identifying the correct echocardiogram video views. Several studies have used deep learning and traditional image-processing techniques for this task. The authors propose an ablation study on a multi-task learning scheme with a hierarchically structured model output that arranges views in a tree structure. The proposed model, named “Multi-task Residual Neural Network (MTRNN) with masked loss”, uses a conditional probabilistic training method and demonstrates superior performance for echocardiogram view classification. While the model has only been validated for the echocardiogram video classification task, it can be easily generalized to any medical image classification scenario with a hierarchical structure among the data labels.