Grant: CIHR Fellowship
Direct Payment
$100,000.00
Respiratory Motion-Embedded Deep Learning Model for Local Failure Prediction and Individualized Radiation Dose Prescription
EF 2021-22
Résumé
The emergence of conformal, high-dose radiotherapy has significantly changed how we treat patients with early-stage lung cancer. Even though progress has been made, studies still show that certain patients are more likely to experience local failure after treatment. Fortunately, research has shown that higher-order image features extracted with a deep learning model can be used to better predict individualized radiation dose, lowering the chance of local failure to below 5% for lung cancer patients. However, the interaction between treatment types, tumor differences, and patient-specific variations makes prediction accuracy complicated. For lung patients, breathing can cause irregular tumor motion, deforming the shape of the tumor and changing its intensity on CT images. Because radiomics features extracted from CT images are affected by motion, we propose to assess the impact of tumor motion on the stability of radiomics features and the prediction accuracy of local failure. We will study free-breathing and motion-tagged 4DCT images from 300 patients, analyzing only regions that have been previously treated with radiation. A computer vision algorithm will be used to automatically track 3D motion of tumors within the region-of-interest. We will examine the motion magnitude, volume deformation, and the stability of features at different breathing phases, and will determine their correlation to the predictability of local failure. We expect this study to help identify the range of motion and the phase of the breathing cycle where features will be most reproducible, and where the prediction of local failure is most accurate. The ability to incorporate motion effects and feature stability into the deep learning model for local failure prediction is key to the clinical realization of personalizing high dose radiation to the lung. While this work is being pioneered by our department, it will help advance radiotherapy treatment both in Canada and elsewhere.
Programme
CIHR Fellowship
Bénéficiaire
Emplacement
Illinois, United States of America
Type
Grant
Source
CIHR Funding Decisions Database
Tel que publié
Id
454424
Name
Teo Peng T
Region
Illinois
Country
United States of America
Orgtype
Research Institution
Instname2
Cancer Research
Theme2
Biomedical
Prcname2
Fellowships - Post-PhD
Projecttitle
Respiratory Motion-Embedded Deep Learning Model for Local Failure Prediction and Individualized Radiation Dose Prescription
Programtype2
Fellowship Programs
Orgnameinp2
Direct Payment
Cihramount2
$100,000
Cihrequipment2
$0
Programname2
CIHR Fellowship
Keyworddelim
Computer Vision; Deep Learning; Local Recurrence; Local Treatment Failure; Lung Cancer; Lung Tumor Motion Tracking; Radiation Oncology; Respiratory Induced Motion; Stereotactic Body Radiotherapy; Treatment Outcome Prediciton
Abstract
The emergence of conformal, high-dose radiotherapy has significantly changed how we treat patients with early-stage lung cancer. Even though progress has been made, studies still show that certain patients are more likely to experience local failure after treatment. Fortunately, research has shown that higher-order image features extracted with a deep learning model can be used to better predict individualized radiation dose, lowering the chance of local failure to below 5% for lung cancer patients. However, the interaction between treatment types, tumor differences, and patient-specific variations makes prediction accuracy complicated. For lung patients, breathing can cause irregular tumor motion, deforming the shape of the tumor and changing its intensity on CT images. Because radiomics features extracted from CT images are affected by motion, we propose to assess the impact of tumor motion on the stability of radiomics features and the prediction accuracy of local failure. We will study free-breathing and motion-tagged 4DCT images from 300 patients, analyzing only regions that have been previously treated with radiation. A computer vision algorithm will be used to automatically track 3D motion of tumors within the region-of-interest. We will examine the motion magnitude, volume deformation, and the stability of features at different breathing phases, and will determine their correlation to the predictability of local failure. We expect this study to help identify the range of motion and the phase of the breathing cycle where features will be most reproducible, and where the prediction of local failure is most accurate. The ability to incorporate motion effects and feature stability into the deep learning model for local failure prediction is key to the clinical realization of personalizing high dose radiation to the lung. While this work is being pioneered by our department, it will help advance radiotherapy treatment both in Canada and elsewhere.
Competitiondate
202110
Approvedterm2
2 yrs 0 mth
Pinamesdelim
Teo, Peng T
Approvedterm
2 ans 0 mois
Cihrcontribution2
~ $100,000~ $0~