Session: Machine Learning, AI and Automation - The Cellar
June 23, 2022 from 11:00am EST to 12:00pm EST
Scientific Session 2 – Machine Learning, AI and Automation
Thursday, June 23, 2022, 11:00-12:00
Scientific Session 2: Machine Learning, AI and Automation – Presentation 1
A Left Hippocampus Sparing Model for Glioblastoma by Utilizing Knowledge-Based Planning and Multi-Criteria Optimization
Shima Yaghoobpour Tari, Amr Heikal, Fan Yang, Deepak Dinakaran, Samir Patel
Cross Cancer Institute
Purpose: A prior study conducted in our institution shows correlation of higher left hippocampal mean dose with neuro-cognitive decline post-radiotherapy for patients with glioblastoma. RapidPlanTM (RP) and Multi-Criteria Optimization (MCO) were used to create a left hippocampus sparing model to improve plans for the glioblastoma patients.
Methods: 97 VMAT plans with conventional (39) and hypofractionated (58) prescriptions were trained to create an RP model called RPv1. 13 independent plans were used for the model validation. All 110 plans were replanned using RPv1 and further optimized using MCO (RPv1+MCO) to improve the mean dose to the left hippocampus while keeping the PTV coverage and OAR constraints clinically acceptable. A new model called RPv2 was created using the improved plans. The RPv2 was validated and all plans were replanned using the RPv2.
Results: The final model, RPv2, has improved the mean dose to the left hippocampus by 50%. The mean dose to the left hippocampus has decreased from 23.21 Gy for the clinically-treated plan to 13.33 Gy for the RPv1+MCO plans and 12.06 Gy for the RPv2 plans. The V95% for PTV was 99.30%±1.49% for the original plan, 99.07%±1.55% for the RPv1+MCO plans, and 99.17%±1.34% for the RPv2 plans. Dose to 1% of the brainstem was also improved on the average from 34.63 Gy for the clinically-treated plan to 28.06 Gy for the RPv2 plans.
Conclusions: A model was developed using RP and MCO that improves sparing the left hippocampus for the glioblastoma patients while maintaining target coverage and OAR sparing to the original clinically-treated plans.
Scientific Session 2: Machine Learning, AI and Automation – Presentation 2
Do dosiomics features derived from planned dose distribution improve biochemical outcome prediction?
Lingyue Sun, Ben Burke, Harvey Quon, Alec Swallow, Charles Kirkby, Wendy Smith
University of Calgary, University of Alberta
Purpose: To investigate whether adding dosiomics features can improve biochemical failure-free survival prediction (BFFS) for prostate EBRT patients.
Methods: This study retrospectively included 1810 patients diagnosed between 2010 to 2016 and treated with curative EBRT at one of the four cancer centers in XX province. 1524 patients from two centers were used for model training and 286 patients from another two centers were used for validation. Three random survival forest models that included different features were built. Model A only included clinical features. Model B included clinical features as well as equivalent uniform dose and tumor control probability derived from dose-volume histograms (DVHs) of the high dose PTV (PTV_High). Model C considered clinical features and 2074 dosiomics features derived from the planned dose distribution for the PTV_High and the CTV_High. Further feature selection was performed to select prognostic features. The three models’ performances were evaluated using Harrell’s concordance index (c-index) and calibration plots for both the training and validation datasets.
Results: Model C selected nine dosiomics features to be prognostic. The c-index of the full training dataset (out-of-bag samples) was 0.72 (0.65), 0.70 (0.65), and 0.74 (0.67) for model A, B, and C, respectively. The c-index of the validation dataset for models A, B, and C was 0.65, 0.64, and 0.66, respectively. The calibration plots for all three models were overall reasonable.
Conclusions: The inclusion of prognostic dosiomics features resulted in an increase of the c-index, suggesting that they can play a role in improving outcome modeling.
Scientific Session 2: Machine Learning, AI and Automation – Presentation 3
Machine-Learning and Texture Analysis of Hyperpolarized 3He MRI Ventilation Predicts Quality-of-life Worsening in Ex-smokers with and without COPD
Maksym Sharma, Marrissa McIntosh, Harkiran Kooner, David McCormack, Grace Parraga
Purpose: To generate an image-guided method for predicting clinically-relevant quality-of-life worsening in ex-smokers with and without chronic obstructive pulmonary disease (COPD).
Methods: Ex-smokers with and without COPD diagnosis (n=50) performed spirometry, 3He magnetic-resonance-imaging (MRI) and computed tomography (CT) at baseline and follow-up (2.4±0.8 years). Participants were dichotomized based on the SGRQ minimal-clinically-important-difference (MCID) (≥4 units). Ventilation-defect-percent (VDP) was quantified and MRI texture features were extracted from gray level run-length, gap-length, size-zone, dependence and co-occurrence-matrices using an open-source PyRadiomics platform. Principle Component Analysis and Boruta analysis that is based on a random-forest algorithm was implemented to select MRI texture features and predictors that independently contribute to the accuracy of machine-learning classification models. Performance was evaluated using the receiver operator characteristic curve, as well as sensitivity and specificity.
Results: We identified six texture feature contributors, which outperformed standard imaging and clinical variables, with the top machine-learning model achieving a classification accuracy of 80.2% at predicting MCID worsening in quality-of-life within 2-3 years. Extracted texture features showed significant moderate (r=0.3-0.7) correlations with VDP, forced expiratory volume in one second (FEV1), ratio of FEV1 to forced-vital-capacity (FEV1/FVC), and the change in SGRQ. The longitudinal change in the Wavelet-Low-High-Pass-Correlation texture feature correlated with the change in VDP (r=-.282, p=.048).
Conclusions: For the first time, 3He MRI ventilation features that predicted MCID quality-of-life worsening in ex-smokers were identified. These pilot results suggest that 3He MRI texture features may provide additional prognostic information to predict clinically-relevant outcomes in ex-smokers.
Scientific Session 2: Machine Learning, AI and Automation – Presentation 4
A Machine Learning Model Predicting Lung Dose to Inform Respiratory Management for Patients Receiving Right-Sided Breast Radiotherapy
Fletcher Barrett, Sarah Quirk, Michael Roumeliotis, Kailyn Stenhouse, Karen Long, Sangjune Lee, Philip McGeachy
University of Calgary - Tom Baker Cancer Centre
Purpose: To develop an anatomy-based lung V20Gy[%] prediction model for right-sided breast cancer patients to inform the need for deep-inspiration breath-hold over free-breathing at the time of CT-SIM.
Methods: This retrospective study included all right-sided breast cancer patients that received: a free-breathing CT scan, radiation treatment between 2015 and 2020, and regional nodal irradiation (n=174). Measurements to predict the total volume of the lung and the volume of the lung in the two treatment field components (tangent and supraclavicular) were: lung width (Anterior-Posterior), lung length (Superior-Inferior), posterior edge of tangent field to distal point of lung, and inferior edge of clavicle to superior point of lung. Three Linear Regression (LR) machine learning models were used to correlate these measurements with lung volume and the volume of lung in the treatment fields, V20Gy, to predict lung V20Gy[%]=V20Gy/lung volume. The data was split into a 70:20:10 ratio for training, validation, and testing. Each LR model was evaluated using 100 random samplings of training and validation data. Using the testing data, performance of each LR model is reported as the average percent difference between true and predicted volumes, and the absolute difference between true and predicted V20Gy[%].
Results: Each LR model’s predictions were within: 9±7% for total lung volume, 17±11% for tangential-field lung volume, and 17±14% for supraclavicular-field lung volume. The resulting calculated V20Gy[%] was within an absolute value of 5±3%.
Conclusions: Predictions of V20Gy[%] were within 5%, on average. Further work will improve predictions to better inform the need for deep-inspiration breath-hold.
Scientific Session 2: Machine Learning, AI and Automation – Presentation 5
Development of a machine learning model to predict dose conformity for external beam partial breast irradiation
Amy Frederick, Alexandra Guebert, Petra Grendarova, Michael Roumeliotis, Sarah Quirk
University of Calgary - Tom Baker Cancer Centre, University of Calgary - Grande Prairie Cancer Centre
Purpose: To predict dose conformity for external beam accelerated partial breast irradiation (EB-APBI) by developing and evaluating a machine learning model.
Methods: A cohort of 275 EB-APBI patients, treated between 2016 and 2019, were included in this study. The dose conformity index (CI), defined as the 95% isodose volume divided by the planning target volume (PTV), and 20 patient-specific and treatment plan characteristics were extracted from each treatment plan. The patient cohort was split into training/validation (80%/20%) datasets. Feature selection using extra-trees, mutual information, and linear regression removed uninformative features. Nested five-fold cross-validation with a grid search was performed over 100 iterations to optimize hyperparameters and estimate performance of a support vector regression (SVR) algorithm. The final model was re-fit to the training dataset and its performance was evaluated on the validation dataset using the mean absolute error (MAE).
Results: The ratio of the dose evaluation volume (DEV, defined as the PTV cropped from the chest wall and skin) to the PTV and the distance between the target and chest wall/skin were the most informative features. In the cross-validation process, the mean (standard deviation) MAE for the training and testing sets were 0.062 (0.004) and 0.075 (0.002), respectively. The final SVR model predicted the CI with a MAE of 0.074. The CI was predicted within ±0.1, ±0.2, and ±0.3 for 73%, 93%, and 100% of samples, respectively.
Conclusions: This predictive model could be used to estimate the achievable CI for EB-APBI planning, refine patient selection, and better understand patient outcomes.
Scientific Session 2: Machine Learning, AI and Automation – Presentation 6
Independent evaluation of RayStation’s deep learning autosegmentation model for structures in the male pelvis
Ming Liu, Dal Granville, Byron Wilson
The Ottawa Hospital Cancer Centre
Purpose: To quantitatively evaluate the performance of a commercial deep learning model for auto-segmenting structures in CT images of the male pelvis.
Methods: We evaluated the male pelvis deep learning auto-segmentation model in RayStation 10A (RaySearch Laboratories, Stockholm, Sweden). This model is based on the 3D UNET architecture and was trained on datasets that are entirely independent from our institution. It performs auto-segmentation of the prostate, bladder, rectum, and femurs. We evaluated its performance by comparing auto-segmented structures to peer-reviewed, manually segmented structures from 100 randomly selected CT image sets accrued in our clinical practice. We quantified the agreement between structures using the dice similarity coefficient (DSC), volume difference, and distance-to-agreement metrics. Three image sets with significant artifacts resulting from hip prostheses were excluded from the study.
Results: DSC values (mean ± standard deviation) were 0.95 ± 0.03, 0.85 ± 0.06, 0.82 ± 0.06, 0.92 ± 0.03, and 0.91 ± 0.03 for bladder, prostate, rectum, left femur, and right femur, respectively. The average distance-to-agreement was within 3.3 mm for all structures. The average volume differences of femurs and rectum were 12 cc and 9 cc, respectively, which likely resulted from variability in defining the superior/inferior extents of these structures. The average time for automatic segmentation was approximately 60 seconds per CT image set.
Conclusions: The agreement between auto-segmented structures and manually segmented structures was similar to previously reported values of interobserver variability. This deep learning model has the potential to reduce segmentation workload in radiotherapy treatment planning workflows.
Scientific Session 2: Machine Learning, AI and Automation – Presentation 7
Deep learning-based autocontouring algorithm for non-invasive intrafractional tumour-tracked radiotherapy (nifteRT) on Linac-MR
Gawon Han, Keith Wachowicz, Nawaid Usmani, Don Yee, Jordan Wong, Gino Fallone, Jihyun Yun
University of Alberta, BC Cancer
Purpose: To develop a neural network-based tumour autocontouring algorithm with implementation of patient- and site-specific hyperparameter optimization (HPO), and to validate its contouring accuracy using in-vivo MR images of liver and prostate cancer patients.
Methods: 2D MR images were acquired at 4 fps using 3T MRI from 6 liver and 23 prostate cancer patients. A deep neural network, U-Net, was applied for autocontouring, and was further improved by implementing HPO using Covariance Matrix Adaptation Evolution Strategy. Six hyperparameters were optimized for each patient, for which the MR images and manual contours were input into the optimization algorithm to find the optimal set of hyperparameters. The U-Net, modified according to the optimized hyperparameters, was subsequently verified by autocontouring the tumour in 70 consecutive images per patient. To evaluate the algorithm’s autocontours, Dice’s coefficient (DC), centroid difference (CD), and Hausdorff distance (HD) were computed between the manual and autocontours. To enhance the execution time of HPO and training of the algorithm, three GPUs were utilized in parallel.
Results: The algorithm was able to perform HPO for each patient and thereby generated tumour autocontours with accuracy comparable to that of experts. The mean (standard deviation) of DC, CD, and HD of twenty-nine patients were 0.92(0.03), 1.47(0.81) mm, and 4.19(1.30) mm, respectively.
Conclusions: To perform non-invasive intrafractional tumour-tracked radiotherapy on Linac-MR, an autocontouring scheme has been developed by implementing HPO to a deep learning-based autocontouring algorithm. The developed algorithm performs patient- and site-specific HPO enabling accurate tumour delineation comparable to that of experts.