Deep learning pour la segmentation des cavités cardiaques en vue de la stratification d’embolie pulmonaire

Details

Supervisors

Macq, Benoît

Faculty

Ecole polytechnique de Louvain

Degree label

Master [120] : ingénieur civil biomédical, à finalité spécialisée

Abstract

enIn Belgium, over 6,000 people suffer from pulmonary embolism every year, and almost 400 of them die. In France, over 47,000 cases are recorded every year. Pulmonary embolism is one of the main causes of cardiovascular mortality, and a major health problem worldwide. The short-term effects can be dangerous if the patient is not properly treated when an embolism is detected. Pulmonary embolisms also have an impact on the heart. In order to identify the risks correctly, different biomarkers such as the volume or diameter of the heart chambers could be used. To calculate these values, a segmentation of the different chambers of the heart is necessary. This requires considerable time on the part of physicians. In the first part of this work, we aim to use artificial intelligence, and more specifically deep learning, to automatically segment the heart’s chambers. The model we have chosen to use is a three-dimensional U-Net. It is trained using a database of public data from the Multi Modality Whole Heart Segmentation challenge, taken by CT scan. These data are twenty in number, with a resolution of 512 × 512 × C, where the value of C varies according to the patient. During training, different resolutions are used and different parameters are tested. Results are quantified using the dice metric. The best dice obtained is 0.72 on average over the different parts of the heart with a resolution of 64 × 64 × 64. A dice of 0.7 is achieved with a resolution of 128 × 128 × 128 and the use of a pre-trained model. In the second part of this work, we received 48 data from cliniques universitaires Saint Luc to evaluate the performance of our model. These data also had a resolution of 512 × 512 × C but had to be modified to be similar to the training data. Finally, the results obtained seem promising, but our model still needs some improvement.