Automatic Delineation of Primary Tumor Volumes from MR Images in Head and Neck Cancer Patients with Deep Learning

Details

Supervisors

Sterpin, Edmond ; Lee, John Aldo

Faculty

Faculté des sciences

Degree label

Master [120] en sciences physiques, à finalité spécialisée: physique médicale

Abstract

Lately, Deep Learning techniques have shown promising results on performing various tasks related to the medical field. More specifically, Convolutional Neural Networks have been used to try to automate tasks linked to the radiation therapy workflow. The latter includes the precise delineation of the tumor and the organs at risk, the dose prescription, orientation of the beams, and so on. Radiotherapy is involved in about half of cancer cases and has to be accurate in order to destroy all the cancer cells while sparing the healthy tissues and organs. In order to achieve this goal, treatment is carefully planned for each patient. During the course of treatment, changes might arise in the patient’s anatomy. Thus, the treatment plan might not be adapted to this new anatomy anymore. Adaptive Radiotherapy addresses this problem by adapting the treatment plan to these changes on the fly. To change the plan so rapidly, the delineation task has to be automated and fast. Convolutional Neural Networks are one candidate for this task. The aim of this thesis is to train a state-of-the-art Convolutional Neural Network (nnU-Net) to automatically delineate the Primary Tumor on Magnetic Resonance Images of 20 Head and Neck Cancer patients to assess how close to a physician’s contours the network can get. Provided the database size, the 3D version of nnU-Net showed acceptable results with mean Dice, Hausdorff Distance and Surface Dice of respectively 0.61 ± 0.21, 16.52 ± 4.25 mm and 0.56 ± 0.23 and median of 0.65, 16.12 mm and 0.56. The network inference on 7 patients showed clinically acceptable results, especially for the lower and posterior part of the Primary Tumor. In addition, an adaptation of nnU-Net is also proposed. This new network, called nnX-Net, allows taking advantage of different imaging modalities even if they are not registered while keeping the advantages offered by nnU-Net.