Session 2A: Analytics and Big Data - Lombardy/Umbria
le 6 juin 2024 from 15h00 CDT to 16h00 CDT
Scientific Session 2 – Analytics and Big Data
Thursday, June 6, 2024, 15:00-16:00
Scientific Session 2A: Analytics and Big Data – Presentation 1
Anonymization by Defacing of Head and Neck Cancer Imaging Datasets for Radiotherapy
Kayla O'Sullivan-Steben, John Kildea
McGill University
Purpose: The increase in public medical imaging datasets has raised concerns about patient reidentification from head CT scans. However, existing defacing tools overlook the need for preservation of critical radiotherapy structures. Therefore, we developed a novel automated anonymization technique that preserves Organs at Risk (OARs) and Planning Target Volumes (PTVs) while removing identifiable features from Head and Neck Cancer (HNC) CTs, CBCTs, and DICOM-RT data.
Methods: Using the eye contours as a landmark, all CT pixels above the inferior-most slice of the eye contour and anterior to the midpoint of the eye were removed. Pixels within PTVs were kept if they intersected with the removed region. The body contour was adjusted to reflect the modified image, and the dose map was anonymized by removing dose voxels corresponding to deleted image pixels. CBCTs were defaced following registration to the associated planning CTs.
Results: Our defacing algorithm successfully anonymized HNC CT scans by removing the eye and forehead regions, while maintaining the integrity of OARs and PTVs. When tested on our in-house dataset of 500 HNC patients with 673 CTs and 7,054 CBCTs, 87.7% of the PTVs were entirely inferior to the cropped region, indicating that most beam entry points were unaffected by the anonymization. Only 4.6% of the PTVs extended beyond the initial cropped region.
Conclusions: We developed a novel defacing algorithm that anonymizes HNC CT scans and related DICOM-RT data while preserving essential structures for radiotherapy research, facilitating the sharing of HNC imaging datasets for Big Data and AI.
Scientific Session 2A: Analytics and Big Data – Presentation 2
Development of geometrical parameters for characterizing anatomical alterations in head and neck cancer patients and evaluating radiotherapy replanning.
Odette Rios-Ibacache, James, Manalad, Aixa X., Andrade Hernandez, Kayla O’Sullivan-Steben, Emily Poon, Luc Galarneau, John Kildea
McGill University, University of Texas Southwestern
Purpose: Head and Neck (H&N) cancer patients undergoing radiotherapy (RT) may experience significant anatomical changes, impacting the effectiveness of the initial treatment plan. Replanning increases clinical staff time, and currently, there is no established method for determining the amount of anatomical variation in the treated region to decide on replanning. This study aims to develop standardized parameters to describe anatomical alterations and influence replanning decisions.
Methods: This retrospective study included 120 H&N patients treated from 2017-2023, with prescribed doses of 60 Gy, 66 Gy, and 70 Gy. We defined over 20 geometrical parameters based on RT structures' shape, such as the Chamfer distance, to describe body shrinkage. Our study included clinical information such as p16 status and smoking history. The Mann-Whitney U test was used to compare replanned and non-replanned patient values distribution, and the Kruskal-Wallis test to evaluate the amount of variation (Δ%) of the parameter and clinical information. In addition, we assessed their rate of change across fractions.
Results: We found significant differences in parameters like body volume and body-to-treatment mask distance between replanned and non-replanned patients, starting as early as fraction 6, Furthermore, we noticed that positive p16 and replanned patients showed a higher Δ%. These parameters have the potential to characterize anatomical changes over time, guide replanning decisions, and prospectively be combined in a machine learning model.
Conclusions: The combination of RT structures and clinical data can be used to create an RT replan predictor, enhancing patient outcomes and streamlining clinical workflows.
Scientific Session 2A: Analytics and Big Data – Presentation 3
Modelling intra-fraction motion and delivery errors in spine SBRT using log files.
Meaghen Shiha, Emily Heath, Eric Vandervoort
Carleton University, The Ottawa Hospital
Purpose: To quantify and examine the impact of patient vertebral target motion during VMAT treatments for spinal metastases.
Methods: Motion log files from CyberKnife treatments were analyzed to evaluate patient motion during the first 5-6 minutes of uninterrupted spine SBRT treatments without the use of rigid immobilization. This duration was chosen as it is representative of typical VMAT beam times. The effect of this motion in a VMAT treatment was simulated using a validated model of 6MV Elekta Synergy FFF beam in BEAMnrc/DOSXYZnrc. Patient motion was synchronized with the delivery log file and motion was modeled as a series of continuous isocenter shifts.
Results: 104 motion intervals were extracted from 24 treatment courses. The mean absolute displacement in each interval along (S/I), (L/R) and (A/P) directions was 0.27 ± 0.28 mm, 0.42 ± 0.38 mm and 0.24 ± 0.2 mm. For our proof in principle, where the motion trace with the maximal mean motion was used, the D90% for the CTV decreased by 5% compared to the plan. A small decrease of 3% in the D0.03cc for the cauda equina was also noted.
Conclusion: We have shown that on average, in the short intervals typical of a VMAT treatment the motion observed is below 1 mm in all directions. The use of a Monte Carlo based framework to simulate the effect of this motion was also demonstrated and showed that when considering patient motion, changes to target coverage of 5% may be observed.
Scientific Session 2A: Analytics and Big Data – Presentation 4
Metabolic Profiling of Murine Radiation-Induced Lung Injury with Raman Spectroscopy and Comparative Machine Learning.
Mitchell Wiebe, Kirsty Milligan, Joan Brewer, Alejandra M. Fuentes, Ramie Ali-Adeeb, Alexandre Brolo, Julian J. Lum, Jeffrey L. Andrews, Christina Haston, Andrew Jirasek
UBC Okanagan, University of Victoria,
Purpose: Radiation-induced lung injury (RILI) is a dose-limiting toxicity for cancer patients receiving thoracic radiotherapy. As such, it is important to characterize metabolic associations with the early and late stages of RILI, namely pneumonitis and pulmonary fibrosis. This work harnesses Raman spectroscopy and supervised machine learning to investigate metabolic associations with radiation pneumonitis and fibrosis in a mouse model.
Methods: Raman spectra were collected from lung tissues of irradiated/non-irradiated mice and labelled as normal, pneumonitis, or fibrosis, based on histological assessment. Spectra were decomposed into metabolic scores via group and basis restricted non-negative matrix factorization, classified with random forest (GBR-NMF-RF), and metabolites predictive of RILI were identified. For comparative context, spectra were decomposed and classified via principal component analysis with random forest, and full spectra were classified with a convolutional neural network (CNN) and logistic regression.
Results: Each methodology was comparable by measure of accuracy and log-loss (all p>0.10 by Mann-Whitney U test), and no methodology was dominant across all classification tasks by measure of area under the receiver operating characteristic curve. GBR-NMF-RF was directly interpretable and identified collagen and specific collagen precursors as top fibrosis predictors, while metabolites with immune and inflammatory functions, such as serine and histidine, were top pneumonitis predictors. The CNN, through heatmap interpretation methods, revealed spectral regions consistent with these metabolites, supporting their associations with RILI.
Conclusions: This work identified RILI-associated metabolites via Raman spectroscopy and supervised machine learning. GBR-NMF-RF and CNN methods showed overlap in metabolite identification and classification performance.
Scientific Session 2A: Analytics and Big Data – Presentation 5
Comparing multi-omic features extracted from two deformable image registration workflows.
Owen Paetkau, Ekaterina Tchistiakova, Charles Kirkby
University of Calgary
Purpose: To identify stable multi-omic features across two validated deformable image registration workflows extracted from head and neck synthetic CTs.
Methods: Deformable image registration workflows were compared using two commercially available software packages (MIM Maestro, Velocity). Workflows to develop head and neck synthetic CT images from the planning CT and weekly cone-beam CT were validated on head and neck patients using contour-based validation. Multi-omic (radiomic and dosiomic) features were extracted from a head and neck patient cohort treated with curative intent radiotherapy (n=64). Feature stability across the two workflows was examined for region-of-interest based feature extraction with original and wavelet filters. Radiomic features were extracted from the pharyngeal constrictors and dosiomic features were extracted from the pharyngeal constrictors and 80% isodose structure. Multi-omic features were compared for stability using an intraclass correlation coefficient with a threshold of >0.80 considered a stable feature between the deformable image registration algorithms.
Results: The pharyngeal constrictor volume and mean dose were found to be statistically similar across the two deformable image registration workflows. Multi-omics features extracted from the pharyngeal constrictor had more stable dosiomic features (20.0%) compared to radiomic features (1.1%). In comparison, features extracted from the 80% isodose line showed more stable dosiomic features (39.5%). Most stable features appeared in the shape and first-order feature classes.
Conclusions: Separately validated deformable image registration workflows produce synthetic CT images with different multi-omic features, with more stable dosiomic features. Care must be taken when using data processed with different workflows in a single model.
Scientific Session 2A: Analytics and Big Data – Presentation 6
Regional analysis of deep-learning texture features for PRISm classification within the CanCOLD study.
Leila Lukhumaidze, James C. Hogg, Jean Bourbeau, Wan C. Tan, Miranda Kirby
Toronto Metropolitan University, University of British Columbia, McGill University Health Centre
Purpose: Previous investigations show individuals with Preserved Ratio Impaired Spirometry (PRISm) have higher ground-glass and reticulation regions in the lung compared to normal-spirometry or COPD individuals, and lower emphysema compared to COPD. However, the regional distribution of fibrosis-like, emphysematous, and normal lung tissue within individuals with PRISm remains understudied. Our objective is to explore the predictive capability of regional quantitative CT texture features in classifying PRISm.
Methods: Stable participants, defined as those that did not change spirometry classification (no-COPD, PRISm, COPD) over three years, from the Canadian Cohort Obstructive Lung Disease (CanCOLD) study were identified. Case-control groups (20 normal spirometry and 20 COPD) were matched to 22 PRISm participants based on demographic and clinical factors. Deep learning-based texture features (%GroundGlass+Reticulation, %Emphysema, and %Normal) were extracted from the lung lobular peel and core regions of CT images. Logistic regression models were employed for group comparison.
Results: %Normal and %GroundGlass+Reticulation features were significant predictors for PRISm vs control, particularly in core regions. Regional and whole lung %Emphysema and %GroundGlass+Reticulation were significant predictors for PRISm vs COPD with the highest odds ratio range [1.07-43.05] observed for %GroundGlass+Reticulation extracted from the core region. Middle and upper lung lobes exhibited consistent predictive capability for PRISm classification.
Conclusions: %GroundGlass+Reticulation in middle and upper lung lobes consistently classified PRISm from control and COPD participants, with core regions showing higher predictive value. %Emphysema distinguished PRISm from COPD, with lobular peel regions being more predictive. These findings highlight the utility of regional CT imaging features in improving PRISm classification.