Ph. D. Project
Title:
Use of Bayesian Deep Learning for Real-Time Monitoring of Pulmonary Lesions from Mono and Stereoscopic Radiological Imaging During Radiotherapy Treatments
Dates:
2024/10/01 - 2027/09/30
Supervisor(s):
Description:
External radiotherapy uses a high-energy X-ray beam, produced by a linear particle accelerator (LINAC), to target tumors from outside the body. Stereotactic body radiotherapy (SBRT) of the lungs is a curative treatment, delivering a high dose of radiation in just a few fractions, precisely targeting small moving pulmonary lesions. This treatment is often performed using volumetric modulated arc therapy (VMAT), which allows for a high dose to be delivered to the target while sparing the surrounding organs at risk (OAR) by rotating the LINAC around the patient and modulating the intensity of the X-rays during delivery.
Since lung tumors move with respiration, this movement must be considered when defining the treatment volume. Generally, this movement is determined from a sample of a few respiratory cycles using 4DCT scanning, but these periods do not always accurately represent breathing during treatment. Therefore, it is crucial to monitor the position of lung tumors during treatment to ensure they receive the intended dose. Reliable position verification can also allow for a reduction in safety margins around the target, thereby limiting irradiation of healthy tissues.
Modern LINACs are equipped with an onboard kilovoltage (kV) imaging system, enabling continuous acquisition of planar kV images (2D) for cone-beam computed tomography (CBCT) reconstruction and for imaging during irradiation. This allows for tracking of selected targets using existing equipment. However, despite high image resolution, the soft tissue contrast is often limited, making precise tumor localization difficult. Some LINACs dedicated to stereotactic treatments have, in addition to the aforementioned imaging system, fixed stereoscopic kV imaging systems (such as Brainlab's ExacTrac) that allow verification of the patient's and target's position in 3D using two simultaneous images.
Lung tumor tracking can be done with or without markers. A marker is a fiducial (usually a gold seed or magnetic coil) implanted by an interventional radiologist. This implantation is an invasive, potentially heavy and risky procedure that is not accessible to all patients treated with lung SBRT.
Markerless tumor position tracking has been demonstrated by coupling individual kV images with 2D models generated from previous volumetric CT data using normalized cross-correlation. The obtained 2D positions can be triangulated to determine the 3D location. However, this method has limitations, especially if the tumor is small, has low density, low contrast, is obstructed by surrounding structures (such as the spine), or if the kV images are very noisy.
Recent developments in dual-energy imaging have shown improvements in tumor tracking, but are not yet clinically available.
Currently, there is only one markerless tracking technology called Synchrony, implemented on Accuray's CyberKnife and Tomotherapy devices. No tracking technique is operational on conventional accelerators, which represent the overwhelming majority of clinical devices (~95% of devices).
In this project, we aim to improve existing markerless detection strategies for conventional accelerators using multimodal deep learning to detect tumors on kV images acquired by monoscopic (onboard imager) and/or stereoscopic (ExacTrac system) imaging. This learning would integrate not only anatomical data, temporal data, but also the patient's breathing information. A key objective will be to evaluate the possibility of extending the capabilities of the dynamic ExacTrac system by performing "quasi-continuous" fluoroscopic imaging, in 3 dimensions, focused on the area of interest for real-time monitoring. This will require the development of an ExacTrac image correction when the LINAC components obstruct its imagers (the irradiation head, the onboard X-ray tube, or the flat panel detectors). This correction will be based on deterministic image processing combined with a Bayesian deep learning strategy (integrating previously acquired information). The support device for this study will be a Varian TrueBeam STx accelerator equipped with the Brainlab ExacTrac system.
Since lung tumors move with respiration, this movement must be considered when defining the treatment volume. Generally, this movement is determined from a sample of a few respiratory cycles using 4DCT scanning, but these periods do not always accurately represent breathing during treatment. Therefore, it is crucial to monitor the position of lung tumors during treatment to ensure they receive the intended dose. Reliable position verification can also allow for a reduction in safety margins around the target, thereby limiting irradiation of healthy tissues.
Modern LINACs are equipped with an onboard kilovoltage (kV) imaging system, enabling continuous acquisition of planar kV images (2D) for cone-beam computed tomography (CBCT) reconstruction and for imaging during irradiation. This allows for tracking of selected targets using existing equipment. However, despite high image resolution, the soft tissue contrast is often limited, making precise tumor localization difficult. Some LINACs dedicated to stereotactic treatments have, in addition to the aforementioned imaging system, fixed stereoscopic kV imaging systems (such as Brainlab's ExacTrac) that allow verification of the patient's and target's position in 3D using two simultaneous images.
Lung tumor tracking can be done with or without markers. A marker is a fiducial (usually a gold seed or magnetic coil) implanted by an interventional radiologist. This implantation is an invasive, potentially heavy and risky procedure that is not accessible to all patients treated with lung SBRT.
Markerless tumor position tracking has been demonstrated by coupling individual kV images with 2D models generated from previous volumetric CT data using normalized cross-correlation. The obtained 2D positions can be triangulated to determine the 3D location. However, this method has limitations, especially if the tumor is small, has low density, low contrast, is obstructed by surrounding structures (such as the spine), or if the kV images are very noisy.
Recent developments in dual-energy imaging have shown improvements in tumor tracking, but are not yet clinically available.
Currently, there is only one markerless tracking technology called Synchrony, implemented on Accuray's CyberKnife and Tomotherapy devices. No tracking technique is operational on conventional accelerators, which represent the overwhelming majority of clinical devices (~95% of devices).
In this project, we aim to improve existing markerless detection strategies for conventional accelerators using multimodal deep learning to detect tumors on kV images acquired by monoscopic (onboard imager) and/or stereoscopic (ExacTrac system) imaging. This learning would integrate not only anatomical data, temporal data, but also the patient's breathing information. A key objective will be to evaluate the possibility of extending the capabilities of the dynamic ExacTrac system by performing "quasi-continuous" fluoroscopic imaging, in 3 dimensions, focused on the area of interest for real-time monitoring. This will require the development of an ExacTrac image correction when the LINAC components obstruct its imagers (the irradiation head, the onboard X-ray tube, or the flat panel detectors). This correction will be based on deterministic image processing combined with a Bayesian deep learning strategy (integrating previously acquired information). The support device for this study will be a Varian TrueBeam STx accelerator equipped with the Brainlab ExacTrac system.
Keywords:
Cancer Treatment, Radiotherapy, Stereotaxy, Artificial intelligence, Image guidance, Tracking
Conditions:
Duration: 3 ans
Employer : Université de Lorraine
Place : CNRS UMR 7039 CRAN, campus Médecine
Profile required: Master 2 degree validated in one of the following fields: radiation physics, image processing, machine learning.
Employer : Université de Lorraine
Place : CNRS UMR 7039 CRAN, campus Médecine
Profile required: Master 2 degree validated in one of the following fields: radiation physics, image processing, machine learning.
Department(s):
| Biology, Signals and Systems in Cancer and Neuroscience |
Funds:
Doctoral award, Ligue Contre le Cancer