Improving treatment plan prediction in protontherapy with transfer learning

Details

Supervisors

Lee, John Aldo

Faculty

Ecole polytechnique de Louvain

Degree label

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

Abstract

Radiation therapy in general is involved in about half of the cancer treatments. Protontherapy is developing worldwide because the ballistics of protons helps to spare the healthy tissues around tumours better than conventional radiotherapy. Radiation therapies need high quality treatment plans to deploy their potential but these still require several hours of human work to be generated. Most of this time is spent in contouring the target volumes and organs at risk and optimising the treatment machine parameters so that the target volume is well irradiated while minimising the dose to the healthy tissues. This thesis is in line with ongoing research, the challenge of which is to achieve a fast and accurate fully automated planning procedure. This study explores the use of knowledge acquired on radiotherapy to improve the quality of protontherapy dose maps predictions made by a U-Net model through transfer learning. One of the proposed strategies was able to improve the validation accuracy of a model trained with 10 patients by 18% as well as the precision on most of the clinical metrics of interest. It also increased the Dice similarity coefficient by 2.5% in average, 6.3% for high doses and 4.8% for low doses. The model benefiting from transfer learning and trained with 10 protontherapy patients reached performances similar to a model trained with 40 protontherapy patients, without transfer learning. Since building a protontherapy dataset with 40 patients takes about a month, this improvement is highly valuable in the clinical practice to speed up the deployment of automatic planning methods.