04/07/2025
By Derek Tang
Date: Tuesday, April 8, 2025
Time: 4-6 p.m.
Location: This will be a virtual defense via Zoom.
Advisor:
Susu Yan, Ph.D., Department of Radiation Oncology, Massachusetts General Hospital and Harvard Medical School
Committee Members:
Erno Sajo, Ph.D., Department of Physics and Applied Physics, University of Massachusetts Lowell
Caroline Desgranges, Ph.D., Department of Physics and Applied Physics, University of Massachusetts Lowell
Abstract:
Infrastructure expenses and capital costs of systems are the most significant hurdles to expanding global access to proton therapy for cancer treatment. Systemic democratization requires the development of new, cost-effective technologies. One possible approach is the evolution of compact proton systems with fixed or limited-angle beamlines, which rely on image guidance and online adaptive radiotherapy to minimize irradiation of healthy tissue. Cone-beam CT (CBCT) images acquired during treatment setup are prone to scatter and beam-hardening artifacts that result in inaccurate Hounsfield unit (HU) values compared to standard CT, compromising the ability to perform precise dose calculations directly on them. To address this challenge, the use of machine learning (ML) algorithms to create HU-corrected synthetic CTs (sCT) from CBCT images is investigated to optimize adaptive workflows for ensuring safe and effective use in compact proton systems.
ML models have shown great promise in quickly generating sCT images but are prone to hallucinations and can distort anatomical features, leading to errors in proton planning. This work investigates four approaches to improve the training of ML models for sCT generation. The first uses modified loss functions to focus the training of generators along the path of a proton beam, in addition to the regions of interest that are contoured by physicians. The second involves the development of a cycle-consistent multi-task automated segmentation and sCT generation model for adaptive workflow acceleration and geometric constraint enforcement to preserve the structure during image translation. Next, a framework using a proton dose-based loss is proposed to directly optimize the training of sCT generators to improve overall treatment planning outcomes. Finally, a method is developed using encoders and StyleGAN to investigate the use of latent vector manipulation, providing users with more input in AI medical image generation.
Model training, validation, and testing was performed using CBCT and CT images with physician contours from 186 patients treated for prostate cancer on an Elekta Agility linear accelerator. Image quality metrics such as the mean absolute error, root-mean-squared error, and structural similarity index metric were used along with DICE scores to determine segmentation performance. Transfer learning on inter- and intra-institutional datasets was also used to assess generalizability across various imaging systems. sCT images were evaluated dosimetrically for proton radiotherapy using RayStation.
Loss-based optimizations proposed in these studies indicate significant improvements in image quality of sCT as compared to established ML models like CycleGAN, though these lead to marginal improvements in treatment planning outcomes. Furthermore, the proposed multi-task network optimizes adaptive radiotherapy workflows by successfully implementing simultaneous segmentation and unsupervised image translation, reducing the amount of time a patient is on the treatment table to minimize uncertainties in adaptive treatments. Additionally, while the results of preliminary StyleGAN training suggest underperformance relative to CycleGAN, it is paramount to highlight the unprecedented model control latent space manipulation provides users. Going forward, this work can inform further development of reliable and efficient models through more targeted approaches to training for sCT generation and the broader use of ML in radiation oncology for improved patient care.