Alana Thibodeau-Antonacci
Ph.D. Student
Department of Physics
Novel Brachytherapy Technology
Artificial Intelligence
Bio
Alana completed a B.Sc. in Physics at the University of Montreal and a M.Sc. in Medical Physics at McGill University. She is currently pursuing a dual Ph.D. in Physics at McGill University and University of Bordeaux.
Current Projects
Development of a Dynamic-Shield Brachytherapy Applicator for the Treatment of Rectal Cancer
Alana is developing a novel applicator equipped with a dynamically-rotating shield for intensity-modulated endorectal brachytherapy. Treatments with the new applicator will ensure optimal irradiation of the tumor while reducing the dose received by the surrounding healthy tissue. This will improve the curative potential of this treatment while simultaneously reducing toxicity and improving patients’ quality of life.
Deep Learning-Based Automatic Segmentation of Rectal Tumors in Endoscopy Images
The goal of this study is to quantify the inter- and intra-observer variability associated with the task of identifying and delineating rectal pathologies in endoscopy images. This information will guide the development of an unbiased deep learning-based segmentation model for rectal tumors in endoscopy images that will be used clinically to assist physicians during radiotherapy treatment planning.
Publications/Awards
2025
In: Medical Physics, vol. 52, iss. 6, pp. 3541–3556, 2025, ISSN: 2473-4209.
@article{nokey,
title = {Trade-off of different deep learning-based auto-segmentation approaches for treatment planning of pediatric craniospinal irradiation autocontouring of OARs for pediatric CSI},
author = {Alana Thibodeau-Antonacci and Marija Popovic and Ozgur Ates and Chia-Ho Hua and James Schneider and Sonia Skamene and Carolyn Freeman and Shirin Abbasinejad Enger and James Man Git Tsui},
url = {https://aapm.onlinelibrary.wiley.com/doi/10.1002/mp.17782},
doi = {10.1002/mp.17782},
issn = {2473-4209},
year = {2025},
date = {2025-04-01},
journal = {Medical Physics},
volume = {52},
issue = {6},
pages = {3541–3556},
abstract = {Background: As auto-segmentation tools become integral to radiotherapy, more commercial products emerge. However, they may not always suit our needs. One notable example is the use of adult-trained commercial software for the contouring of organs at risk (OARs) of pediatric patients.
Purpose: This study aimed to compare three auto-segmentation approaches in the context of pediatric craniospinal irradiation (CSI): commercial, out-of-the-box, and in-house.
Methods: CT scans from 142 pediatric patients undergoing CSI were obtained from St. Jude Children's Research Hospital (training: 115; validation: 27). A test dataset comprising 16 CT scans was collected from the McGill University Health Centre. All images underwent manual delineation of 18 OARs. LimbusAI v1.7 served as the commercial product, while nnU-Net was trained for benchmarking. Additionally, a two-step in-house approach was pursued where smaller 3D CT scans containing the OAR of interest were first recovered and then used as input to train organ-specific models. Three variants of the U-Net architecture were explored: a basic U-Net, an attention U-Net, and a 2.5D U-Net. The dice similarity coefficient (DSC) assessed segmentation accuracy, and the DSC trend with age was investigated (Mann-Kendall test). A radiation oncologist determined the clinical acceptability of all contours using a five-point Likert scale.
Results: Differences in the contours between the validation and test datasets reflected the distinct institutional standards. The lungs and left kidney displayed an increasing age-related trend of the DSC values with LimbusAI on the validation and test datasets. LimbusAI contours of the esophagus were often truncated distally and mistaken for the trachea for younger patients, resulting in a DSC score of less than 0.5 on both datasets. Additionally, the kidneys frequently exhibited false negatives, leading to mean DSC values that were up to 0.11 lower on the validation set and 0.07 on the test set compared to the other models. Overall, nnU-Net achieved good performance for body organs but exhibited difficulty differentiating the laterality of head structures, resulting in a large variation of DSC values with the standard deviation reaching 0.35 for the lenses. All in-house models generally had similar DSC values when compared against each other and nnU-Net. Inference time on the test data was between 47-55 min on a Central Processing Unit (CPU) for the in-house models, while it was 1h 21m with a V100 Graphics Processing Unit (GPU) for nnU-Net.
Conclusions: LimbusAI could not adapt well to pediatric anatomy for the esophagus and the kidneys. When commercial products do not suit the study population, the nnU-Net is a viable option but requires adjustments. In resource-constrained settings, the in-house model provides an alternative. Implementing an automated segmentation tool requires careful monitoring and quality assurance regardless of the approach.},
keywords = {},
pubstate = {published},
tppubtype = {article}
}
Background: As auto-segmentation tools become integral to radiotherapy, more commercial products emerge. However, they may not always suit our needs. One notable example is the use of adult-trained commercial software for the contouring of organs at risk (OARs) of pediatric patients.
Purpose: This study aimed to compare three auto-segmentation approaches in the context of pediatric craniospinal irradiation (CSI): commercial, out-of-the-box, and in-house.
Methods: CT scans from 142 pediatric patients undergoing CSI were obtained from St. Jude Children's Research Hospital (training: 115; validation: 27). A test dataset comprising 16 CT scans was collected from the McGill University Health Centre. All images underwent manual delineation of 18 OARs. LimbusAI v1.7 served as the commercial product, while nnU-Net was trained for benchmarking. Additionally, a two-step in-house approach was pursued where smaller 3D CT scans containing the OAR of interest were first recovered and then used as input to train organ-specific models. Three variants of the U-Net architecture were explored: a basic U-Net, an attention U-Net, and a 2.5D U-Net. The dice similarity coefficient (DSC) assessed segmentation accuracy, and the DSC trend with age was investigated (Mann-Kendall test). A radiation oncologist determined the clinical acceptability of all contours using a five-point Likert scale.
Results: Differences in the contours between the validation and test datasets reflected the distinct institutional standards. The lungs and left kidney displayed an increasing age-related trend of the DSC values with LimbusAI on the validation and test datasets. LimbusAI contours of the esophagus were often truncated distally and mistaken for the trachea for younger patients, resulting in a DSC score of less than 0.5 on both datasets. Additionally, the kidneys frequently exhibited false negatives, leading to mean DSC values that were up to 0.11 lower on the validation set and 0.07 on the test set compared to the other models. Overall, nnU-Net achieved good performance for body organs but exhibited difficulty differentiating the laterality of head structures, resulting in a large variation of DSC values with the standard deviation reaching 0.35 for the lenses. All in-house models generally had similar DSC values when compared against each other and nnU-Net. Inference time on the test data was between 47-55 min on a Central Processing Unit (CPU) for the in-house models, while it was 1h 21m with a V100 Graphics Processing Unit (GPU) for nnU-Net.
Conclusions: LimbusAI could not adapt well to pediatric anatomy for the esophagus and the kidneys. When commercial products do not suit the study population, the nnU-Net is a viable option but requires adjustments. In resource-constrained settings, the in-house model provides an alternative. Implementing an automated segmentation tool requires careful monitoring and quality assurance regardless of the approach.
Purpose: This study aimed to compare three auto-segmentation approaches in the context of pediatric craniospinal irradiation (CSI): commercial, out-of-the-box, and in-house.
Methods: CT scans from 142 pediatric patients undergoing CSI were obtained from St. Jude Children's Research Hospital (training: 115; validation: 27). A test dataset comprising 16 CT scans was collected from the McGill University Health Centre. All images underwent manual delineation of 18 OARs. LimbusAI v1.7 served as the commercial product, while nnU-Net was trained for benchmarking. Additionally, a two-step in-house approach was pursued where smaller 3D CT scans containing the OAR of interest were first recovered and then used as input to train organ-specific models. Three variants of the U-Net architecture were explored: a basic U-Net, an attention U-Net, and a 2.5D U-Net. The dice similarity coefficient (DSC) assessed segmentation accuracy, and the DSC trend with age was investigated (Mann-Kendall test). A radiation oncologist determined the clinical acceptability of all contours using a five-point Likert scale.
Results: Differences in the contours between the validation and test datasets reflected the distinct institutional standards. The lungs and left kidney displayed an increasing age-related trend of the DSC values with LimbusAI on the validation and test datasets. LimbusAI contours of the esophagus were often truncated distally and mistaken for the trachea for younger patients, resulting in a DSC score of less than 0.5 on both datasets. Additionally, the kidneys frequently exhibited false negatives, leading to mean DSC values that were up to 0.11 lower on the validation set and 0.07 on the test set compared to the other models. Overall, nnU-Net achieved good performance for body organs but exhibited difficulty differentiating the laterality of head structures, resulting in a large variation of DSC values with the standard deviation reaching 0.35 for the lenses. All in-house models generally had similar DSC values when compared against each other and nnU-Net. Inference time on the test data was between 47-55 min on a Central Processing Unit (CPU) for the in-house models, while it was 1h 21m with a V100 Graphics Processing Unit (GPU) for nnU-Net.
Conclusions: LimbusAI could not adapt well to pediatric anatomy for the esophagus and the kidneys. When commercial products do not suit the study population, the nnU-Net is a viable option but requires adjustments. In resource-constrained settings, the in-house model provides an alternative. Implementing an automated segmentation tool requires careful monitoring and quality assurance regardless of the approach.
Dosimetric impact of positional uncertainties and a robust optimization approach for rectal intensity-modulated brachytherapy Journal Article
In: Medical Physics, vol. 52, iss. 6, pp. 3528–3540, 2025, ISSN: 0094-2405.
@article{nokey,
title = {Dosimetric impact of positional uncertainties and a robust optimization approach for rectal intensity-modulated brachytherapy},
author = {Björn Morén and Alana Thibodeau-Antonacci and Jonathan Kalinowski and Shirin A. Enger},
url = {https://aapm.onlinelibrary.wiley.com/doi/10.1002/mp.17800},
doi = {10.1002/mp.17800},
issn = {0094-2405},
year = {2025},
date = {2025-03-31},
journal = {Medical Physics},
volume = {52},
issue = {6},
pages = {3528–3540},
abstract = {Background: Intensity-modulated brachytherapy (IMBT) employs rotating high-Z shields during treatment to decrease radiation in certain directions and conform the dose distribution to the target volume. Prototypes for dynamic IMBT have been proposed for prostate, cervical, and rectal cancer.
Purpose: We considered two shielded applicators for IMBT rectal cancer treatment and investigated how rotational uncertainties in the shield angle and translational uncertainties in the source position affect plan evaluation criteria.
Methods: The effect of rotational errors of 3∘ , 5∘ and 10∘ , and translational errors of 1, 2 and 3 mm on evaluation criteria were investigated for shields with
180
∘
and
90
∘
emission windows. Further, a robust optimization approach based on quadratic penalties that includes scenarios with errors was proposed. The extent to which dosimetric effects of positional errors can be mitigated with this model was evaluated compared to a quadratic penalty model without scenarios with errors. A retrospective rectal cancer data set of ten patients was included in this study. Treatment planning was performed using the Monte Carlo-based treatment planning system, RapidBrachyMCTPS.
Results: For the largest investigated rotational error of
±
10
∘
, the clinical target volume
D
90
remained, on average, within
5
%
of the result without error, while the contralateral healthy rectal wall experienced an increase in the mean
D
0.1
c
c
,
D
2
c
c
, and
D
50
of
26
%
,
9
%
, and
1
%
for the
180
∘
shield and of 32%, 9%, and 2% for the
90
∘
shield. For translational errors of
±
2
mm, there were increases in dosimetric indices for both the superior (sup) and inferior (inf) dose spill regions. Specifically, for the
180
∘
shield, the
D
0.1
c
c
,
D
2
c
c
, and
D
50
increased by
13
%
,
11
%
, and
10
%
, respectively, for the sup region, and by
26
%
,
15
%
, and
11
%
, respectively, for the inf region. Similar results were obtained with the
90
∘
shield. Overall, the robust and traditional models had similar results. However, the number of active dwell positions obtained with the robust model was larger, and the longest dwell time was shorter.
Conclusions: We have quantified the effect of rotational shield and translational source errors of various magnitudes on evaluation criteria for rectal IMBT. The robust optimization approach was generally not able to mitigate positional errors. However, it resulted in more homogeneous dwell times, which can be beneficial in conventional high-dose-rate brachytherapy to avoid hot spots around specific dwell positions.},
keywords = {},
pubstate = {published},
tppubtype = {article}
}
Background: Intensity-modulated brachytherapy (IMBT) employs rotating high-Z shields during treatment to decrease radiation in certain directions and conform the dose distribution to the target volume. Prototypes for dynamic IMBT have been proposed for prostate, cervical, and rectal cancer.
Purpose: We considered two shielded applicators for IMBT rectal cancer treatment and investigated how rotational uncertainties in the shield angle and translational uncertainties in the source position affect plan evaluation criteria.
Methods: The effect of rotational errors of 3∘ , 5∘ and 10∘ , and translational errors of 1, 2 and 3 mm on evaluation criteria were investigated for shields with
180
∘
and
90
∘
emission windows. Further, a robust optimization approach based on quadratic penalties that includes scenarios with errors was proposed. The extent to which dosimetric effects of positional errors can be mitigated with this model was evaluated compared to a quadratic penalty model without scenarios with errors. A retrospective rectal cancer data set of ten patients was included in this study. Treatment planning was performed using the Monte Carlo-based treatment planning system, RapidBrachyMCTPS.
Results: For the largest investigated rotational error of
±
10
∘
, the clinical target volume
D
90
remained, on average, within
5
%
of the result without error, while the contralateral healthy rectal wall experienced an increase in the mean
D
0.1
c
c
,
D
2
c
c
, and
D
50
of
26
%
,
9
%
, and
1
%
for the
180
∘
shield and of 32%, 9%, and 2% for the
90
∘
shield. For translational errors of
±
2
mm, there were increases in dosimetric indices for both the superior (sup) and inferior (inf) dose spill regions. Specifically, for the
180
∘
shield, the
D
0.1
c
c
,
D
2
c
c
, and
D
50
increased by
13
%
,
11
%
, and
10
%
, respectively, for the sup region, and by
26
%
,
15
%
, and
11
%
, respectively, for the inf region. Similar results were obtained with the
90
∘
shield. Overall, the robust and traditional models had similar results. However, the number of active dwell positions obtained with the robust model was larger, and the longest dwell time was shorter.
Conclusions: We have quantified the effect of rotational shield and translational source errors of various magnitudes on evaluation criteria for rectal IMBT. The robust optimization approach was generally not able to mitigate positional errors. However, it resulted in more homogeneous dwell times, which can be beneficial in conventional high-dose-rate brachytherapy to avoid hot spots around specific dwell positions.
Purpose: We considered two shielded applicators for IMBT rectal cancer treatment and investigated how rotational uncertainties in the shield angle and translational uncertainties in the source position affect plan evaluation criteria.
Methods: The effect of rotational errors of 3∘ , 5∘ and 10∘ , and translational errors of 1, 2 and 3 mm on evaluation criteria were investigated for shields with
180
∘
and
90
∘
emission windows. Further, a robust optimization approach based on quadratic penalties that includes scenarios with errors was proposed. The extent to which dosimetric effects of positional errors can be mitigated with this model was evaluated compared to a quadratic penalty model without scenarios with errors. A retrospective rectal cancer data set of ten patients was included in this study. Treatment planning was performed using the Monte Carlo-based treatment planning system, RapidBrachyMCTPS.
Results: For the largest investigated rotational error of
±
10
∘
, the clinical target volume
D
90
remained, on average, within
5
%
of the result without error, while the contralateral healthy rectal wall experienced an increase in the mean
D
0.1
c
c
,
D
2
c
c
, and
D
50
of
26
%
,
9
%
, and
1
%
for the
180
∘
shield and of 32%, 9%, and 2% for the
90
∘
shield. For translational errors of
±
2
mm, there were increases in dosimetric indices for both the superior (sup) and inferior (inf) dose spill regions. Specifically, for the
180
∘
shield, the
D
0.1
c
c
,
D
2
c
c
, and
D
50
increased by
13
%
,
11
%
, and
10
%
, respectively, for the sup region, and by
26
%
,
15
%
, and
11
%
, respectively, for the inf region. Similar results were obtained with the
90
∘
shield. Overall, the robust and traditional models had similar results. However, the number of active dwell positions obtained with the robust model was larger, and the longest dwell time was shorter.
Conclusions: We have quantified the effect of rotational shield and translational source errors of various magnitudes on evaluation criteria for rectal IMBT. The robust optimization approach was generally not able to mitigate positional errors. However, it resulted in more homogeneous dwell times, which can be beneficial in conventional high-dose-rate brachytherapy to avoid hot spots around specific dwell positions.
2022
A121 QUANTIFYING INTER-OBSERVER VARIABILITY IN THE SEGMENTATION OF RECTAL TUMORS IN ENDOSCOPY IMAGES AND ITS EFFECTS ON DEEP LEARNING Journal Article
In: Journal of the Canadian Association of Gastroenterology, vol. 5, no. Supplement_1, pp. 140–142, 2022.
@article{weishaupt2022a121,
title = {A121 QUANTIFYING INTER-OBSERVER VARIABILITY IN THE SEGMENTATION OF RECTAL TUMORS IN ENDOSCOPY IMAGES AND ITS EFFECTS ON DEEP LEARNING},
author = {Luca L Weishaupt and Te Vuong and Alana Thibodeau-Antonacci and A Garant and KS Singh and C Miller and A Martin and Shirin A. Enger},
year = {2022},
date = {2022-01-01},
urldate = {2022-01-01},
journal = {Journal of the Canadian Association of Gastroenterology},
volume = {5},
number = {Supplement_1},
pages = {140--142},
publisher = {Oxford University Press US},
keywords = {},
pubstate = {published},
tppubtype = {article}
}
Development of a Novel MRI-Compatible Applicator for Intensity Modulated Rectal Brachytherapy Proceedings Article
In: MEDICAL PHYSICS, pp. E240–E240, WILEY 111 RIVER ST, HOBOKEN 07030-5774, NJ USA 2022.
@inproceedings{thibodeau2022development,
title = {Development of a Novel MRI-Compatible Applicator for Intensity Modulated Rectal Brachytherapy},
author = {Alana Thibodeau-Antonacci and Te Vuong and B Liontis and F Rayes and S Pande and Shirin A. Enger},
year = {2022},
date = {2022-01-01},
urldate = {2022-01-01},
booktitle = {MEDICAL PHYSICS},
volume = {49},
number = {6},
pages = {E240--E240},
organization = {WILEY 111 RIVER ST, HOBOKEN 07030-5774, NJ USA},
keywords = {},
pubstate = {published},
tppubtype = {inproceedings}
}
Gafchromic film and scintillator detector measurements in phantom with a novel intensity-modulated brachytherapy endorectal shield Proceedings Article
In: MEDICAL PHYSICS, pp. 5688–5689, WILEY 111 RIVER ST, HOBOKEN 07030-5774, NJ USA 2022.
@inproceedings{thibodeau2022gafchromic,
title = {Gafchromic film and scintillator detector measurements in phantom with a novel intensity-modulated brachytherapy endorectal shield},
author = {Alana Thibodeau-Antonacci and Shirin A. Enger and Hamed Bekerat and Te Vuong},
year = {2022},
date = {2022-01-01},
urldate = {2022-01-01},
booktitle = {MEDICAL PHYSICS},
volume = {49},
number = {8},
pages = {5688--5689},
organization = {WILEY 111 RIVER ST, HOBOKEN 07030-5774, NJ USA},
keywords = {},
pubstate = {published},
tppubtype = {inproceedings}
}
PO-1325 Automated rectal tumor segmentation with inter-observer variability-based uncertainty estimates Journal Article
In: Radiotherapy and Oncology, vol. 170, pp. S1120–S1121, 2022.
@article{weishaupt2022po,
title = {PO-1325 Automated rectal tumor segmentation with inter-observer variability-based uncertainty estimates},
author = {Luca L. Weishaupt and Te Vuong and Alana Thibodeau-Antonacci and A Garant and K Singh and C Miller and A Martin and F Schmitt-Ulms and Shirin A. Enger},
year = {2022},
date = {2022-01-01},
urldate = {2022-01-01},
journal = {Radiotherapy and Oncology},
volume = {170},
pages = {S1120--S1121},
publisher = {Elsevier},
keywords = {},
pubstate = {published},
tppubtype = {article}
}
2021
Canada Graduate Scholarship – Doctoral Program award
2021.
@award{Thibodeau-Antonacci2021d,
title = {Canada Graduate Scholarship – Doctoral Program},
author = {Alana Thibodeau-Antonacci},
url = {https://www.nserc-crsng.gc.ca/students-etudiants/pg-cs/cgsd-bescd_eng.asp},
year = {2021},
date = {2021-09-01},
organization = {NSERC},
keywords = {},
pubstate = {published},
tppubtype = {award}
}
Deep learning based tumor segmentation of endoscopy images for rectal cancer patients Presentation
ESTRO Annual meeting, 27.08.2021.
@misc{Weishaupt2021b,
title = {Deep learning based tumor segmentation of endoscopy images for rectal cancer patients},
author = {Luca L. Weishaupt and Alana Thibodeau-Antonacci and Aurelie Garant and Kelita Singh and Corey Miller and Té Vuong and Shirin A. Enger},
url = {https://www.estro.org/Congresses/ESTRO-2021/610/posterdiscussion34-deep-learningforauto-contouring/3710/deeplearning-basedtumorsegmentationofendoscopyimag},
year = {2021},
date = {2021-08-27},
urldate = {2021-08-27},
abstract = {Purpose or Objective
The objective of this study was to develop an automated rectal tumor segmentation algorithm from endoscopy images. The algorithm will be used in a future multimodal treatment outcome prediction model. Currently, treatment outcome prediction models rely on manual segmentations of regions of interest, which are prone to inter-observer variability. To quantify this human error and demonstrate the feasibility of automated endoscopy image segmentation, we compare three deep learning architectures.
Material and Methods
A gastrointestinal physician (G1) segmented 550 endoscopy images of rectal tumors into tumor and non-tumor regions. To quantify the inter-observer variability, a second gastrointestinal physician (G2) contoured 319 of the images independently.
The 550 images and annotations from G1 were divided into 408 training, 82 validation, and 60 testing sets. Three deep learning architectures were trained; a fully convolutional neural network (FCN32), a U-Net, and a SegNet. These architectures have been used for robust medical image segmentation in previous studies.
All models were trained on a CPU supercomputing cluster. Data augmentation in the form of random image transformations, including scaling, rotation, shearing, Gaussian blurring, and noise addition, was used to improve the models' robustness.
The neural networks' output went through a final layer of noise removal and hole filling before evaluation. Finally, the segmentations from G2 and the neural networks' predictions were compared against the ground truth labels from G1.
Results
The FCN32, U-Net, and SegNet had average segmentation times of 0.77, 0.48, and 0.43 seconds per image, respectively. The average segmentation time per image for G1 and G2 were 10 and 8 seconds, respectively.
All the ground truth labels contained tumors, but G2 and the deep learning models did not always find tumors in the images. The scores are based on the agreement of tumor contours with G1’s ground truth and were thus only computed for images in which tumor was found. The automated segmentation algorithms consistently achieved equal or better scores than G2's manual segmentations. G2's low F1/DICE and precision scores indicate poor agreement between the manual contours.
Conclusion
There is a need for robust and accurate segmentation algorithms for rectal tumor segmentation since manual segmentation of these tumors is susceptible to significant inter-observer variability. The deep learning-based segmentation algorithms proposed in this study are more efficient and achieved a higher agreement with our manual ground truth segmentations than a second expert annotator. Future studies will investigate how to train deep learning models on multiple ground truth annotations to prevent learning observer biases.},
howpublished = {ESTRO Annual meeting},
keywords = {},
pubstate = {published},
tppubtype = {presentation}
}
Purpose or Objective
The objective of this study was to develop an automated rectal tumor segmentation algorithm from endoscopy images. The algorithm will be used in a future multimodal treatment outcome prediction model. Currently, treatment outcome prediction models rely on manual segmentations of regions of interest, which are prone to inter-observer variability. To quantify this human error and demonstrate the feasibility of automated endoscopy image segmentation, we compare three deep learning architectures.
Material and Methods
A gastrointestinal physician (G1) segmented 550 endoscopy images of rectal tumors into tumor and non-tumor regions. To quantify the inter-observer variability, a second gastrointestinal physician (G2) contoured 319 of the images independently.
The 550 images and annotations from G1 were divided into 408 training, 82 validation, and 60 testing sets. Three deep learning architectures were trained; a fully convolutional neural network (FCN32), a U-Net, and a SegNet. These architectures have been used for robust medical image segmentation in previous studies.
All models were trained on a CPU supercomputing cluster. Data augmentation in the form of random image transformations, including scaling, rotation, shearing, Gaussian blurring, and noise addition, was used to improve the models' robustness.
The neural networks' output went through a final layer of noise removal and hole filling before evaluation. Finally, the segmentations from G2 and the neural networks' predictions were compared against the ground truth labels from G1.
Results
The FCN32, U-Net, and SegNet had average segmentation times of 0.77, 0.48, and 0.43 seconds per image, respectively. The average segmentation time per image for G1 and G2 were 10 and 8 seconds, respectively.
All the ground truth labels contained tumors, but G2 and the deep learning models did not always find tumors in the images. The scores are based on the agreement of tumor contours with G1’s ground truth and were thus only computed for images in which tumor was found. The automated segmentation algorithms consistently achieved equal or better scores than G2's manual segmentations. G2's low F1/DICE and precision scores indicate poor agreement between the manual contours.
Conclusion
There is a need for robust and accurate segmentation algorithms for rectal tumor segmentation since manual segmentation of these tumors is susceptible to significant inter-observer variability. The deep learning-based segmentation algorithms proposed in this study are more efficient and achieved a higher agreement with our manual ground truth segmentations than a second expert annotator. Future studies will investigate how to train deep learning models on multiple ground truth annotations to prevent learning observer biases.
The objective of this study was to develop an automated rectal tumor segmentation algorithm from endoscopy images. The algorithm will be used in a future multimodal treatment outcome prediction model. Currently, treatment outcome prediction models rely on manual segmentations of regions of interest, which are prone to inter-observer variability. To quantify this human error and demonstrate the feasibility of automated endoscopy image segmentation, we compare three deep learning architectures.
Material and Methods
A gastrointestinal physician (G1) segmented 550 endoscopy images of rectal tumors into tumor and non-tumor regions. To quantify the inter-observer variability, a second gastrointestinal physician (G2) contoured 319 of the images independently.
The 550 images and annotations from G1 were divided into 408 training, 82 validation, and 60 testing sets. Three deep learning architectures were trained; a fully convolutional neural network (FCN32), a U-Net, and a SegNet. These architectures have been used for robust medical image segmentation in previous studies.
All models were trained on a CPU supercomputing cluster. Data augmentation in the form of random image transformations, including scaling, rotation, shearing, Gaussian blurring, and noise addition, was used to improve the models' robustness.
The neural networks' output went through a final layer of noise removal and hole filling before evaluation. Finally, the segmentations from G2 and the neural networks' predictions were compared against the ground truth labels from G1.
Results
The FCN32, U-Net, and SegNet had average segmentation times of 0.77, 0.48, and 0.43 seconds per image, respectively. The average segmentation time per image for G1 and G2 were 10 and 8 seconds, respectively.
All the ground truth labels contained tumors, but G2 and the deep learning models did not always find tumors in the images. The scores are based on the agreement of tumor contours with G1’s ground truth and were thus only computed for images in which tumor was found. The automated segmentation algorithms consistently achieved equal or better scores than G2's manual segmentations. G2's low F1/DICE and precision scores indicate poor agreement between the manual contours.
Conclusion
There is a need for robust and accurate segmentation algorithms for rectal tumor segmentation since manual segmentation of these tumors is susceptible to significant inter-observer variability. The deep learning-based segmentation algorithms proposed in this study are more efficient and achieved a higher agreement with our manual ground truth segmentations than a second expert annotator. Future studies will investigate how to train deep learning models on multiple ground truth annotations to prevent learning observer biases.
Mitacs Globalink Research Award award
2021.
@award{Thibodeau-Antonacci2021c,
title = {Mitacs Globalink Research Award},
author = {Alana Thibodeau-Antonacci and Hossein Jafarzadeh and Liam Carroll and Luca L. Weishaupt},
url = {https://www.mitacs.ca/en/programs/globalink/globalink-research-award},
year = {2021},
date = {2021-07-01},
urldate = {2021-07-01},
organization = {MITACS},
abstract = {The Mitacs Globalink Research Award (GRA) supports research collaborations between Canada and select partner organizations and eligible countries and regions. It was awarded to Alana Thibodeau-Antonacci, Hossein Jafarzadeh, Liam Carroll and Luca L. Weishaupt.
Under the joint supervision of a home and host professor, successful senior undergraduate students, graduate students, as well as postdoctoral fellows will receive a $6,000 research award to conduct a 12- to 24-week research project in the other country. Awards are offered in partnership with Mitacs’s Canadian academic partners (and, in some cases, with Mitacs’s international partners) and are subject to available funding. },
howpublished = {Mitacs},
keywords = {},
pubstate = {published},
tppubtype = {award}
}
The Mitacs Globalink Research Award (GRA) supports research collaborations between Canada and select partner organizations and eligible countries and regions. It was awarded to Alana Thibodeau-Antonacci, Hossein Jafarzadeh, Liam Carroll and Luca L. Weishaupt.
Under the joint supervision of a home and host professor, successful senior undergraduate students, graduate students, as well as postdoctoral fellows will receive a $6,000 research award to conduct a 12- to 24-week research project in the other country. Awards are offered in partnership with Mitacs’s Canadian academic partners (and, in some cases, with Mitacs’s international partners) and are subject to available funding.
Under the joint supervision of a home and host professor, successful senior undergraduate students, graduate students, as well as postdoctoral fellows will receive a $6,000 research award to conduct a 12- to 24-week research project in the other country. Awards are offered in partnership with Mitacs’s Canadian academic partners (and, in some cases, with Mitacs’s international partners) and are subject to available funding.
Inter-Observer Variability and Deep Learning in Rectal Tumor Segmentation from Endoscopy Images Presentation
The COMP Annual Scientific Meeting 2021, 22.06.2021.
@misc{Weishaupt2021c,
title = {Inter-Observer Variability and Deep Learning in Rectal Tumor Segmentation from Endoscopy Images},
author = {Luca L. Weishaupt and Alana Thibodeau-Antonacci and Aurelie Garant and Kelita Singh and Corey Miller and Té Vuong and Shirin A. Enger},
year = {2021},
date = {2021-06-22},
urldate = {2021-06-22},
abstract = {Purpose
To develop an automated rectal tumor segmentation algorithm from endoscopy images.
Material/Methods
A gastrointestinal physician (G1) segmented 2005 endoscopy images into tumor and non-tumor
regions. To quantify the inter-observer variability, a second gastrointestinal physician (G2)
contoured the images independently.
Three deep-learning architectures used for robust medical image segmentation in previous
studies were trained: a fully convolutional neural network (FCN32), a U-Net, and a SegNet.
Since the majority of the images did not contain tumors, two methods were compared for
training. Models were trained using only tumor images (M1) and all images (M2). G1’s images
and annotations were divided into 408 training, 82 validation, and 60 testing sets for M1, 1181
training, 372 validation, and 452 testing sets for M2.
Finally, segmentations from G2 and neural networks' predictions were compared against ground
truth labels from G1, and F1 scores were computed for images where both physicians found
tumors.
Results
The deep-learning segmentation took less than 1 second, while manual segmentation took
approximately 10 seconds per image.
The M1’s models consistently achieved equal or better scores (SegNet F1:0.80±0.08) than G2's
manual segmentations (F1:0.68±0.25). G2's low F1/DICE and precision scores indicate poor
agreement between the manual contours. Models from M2 achieved lower scores than G2 and
M1’s models since they demonstrated a strong bias towards predicting no tumor for all images.
Conclusion
Future studies will investigate training on an equal number of images with/without tumor, using
ground truth contours from multiple experts simultaneously.},
howpublished = {The COMP Annual Scientific Meeting 2021},
keywords = {},
pubstate = {published},
tppubtype = {presentation}
}
Purpose
To develop an automated rectal tumor segmentation algorithm from endoscopy images.
Material/Methods
A gastrointestinal physician (G1) segmented 2005 endoscopy images into tumor and non-tumor
regions. To quantify the inter-observer variability, a second gastrointestinal physician (G2)
contoured the images independently.
Three deep-learning architectures used for robust medical image segmentation in previous
studies were trained: a fully convolutional neural network (FCN32), a U-Net, and a SegNet.
Since the majority of the images did not contain tumors, two methods were compared for
training. Models were trained using only tumor images (M1) and all images (M2). G1’s images
and annotations were divided into 408 training, 82 validation, and 60 testing sets for M1, 1181
training, 372 validation, and 452 testing sets for M2.
Finally, segmentations from G2 and neural networks' predictions were compared against ground
truth labels from G1, and F1 scores were computed for images where both physicians found
tumors.
Results
The deep-learning segmentation took less than 1 second, while manual segmentation took
approximately 10 seconds per image.
The M1’s models consistently achieved equal or better scores (SegNet F1:0.80±0.08) than G2's
manual segmentations (F1:0.68±0.25). G2's low F1/DICE and precision scores indicate poor
agreement between the manual contours. Models from M2 achieved lower scores than G2 and
M1’s models since they demonstrated a strong bias towards predicting no tumor for all images.
Conclusion
Future studies will investigate training on an equal number of images with/without tumor, using
ground truth contours from multiple experts simultaneously.
To develop an automated rectal tumor segmentation algorithm from endoscopy images.
Material/Methods
A gastrointestinal physician (G1) segmented 2005 endoscopy images into tumor and non-tumor
regions. To quantify the inter-observer variability, a second gastrointestinal physician (G2)
contoured the images independently.
Three deep-learning architectures used for robust medical image segmentation in previous
studies were trained: a fully convolutional neural network (FCN32), a U-Net, and a SegNet.
Since the majority of the images did not contain tumors, two methods were compared for
training. Models were trained using only tumor images (M1) and all images (M2). G1’s images
and annotations were divided into 408 training, 82 validation, and 60 testing sets for M1, 1181
training, 372 validation, and 452 testing sets for M2.
Finally, segmentations from G2 and neural networks' predictions were compared against ground
truth labels from G1, and F1 scores were computed for images where both physicians found
tumors.
Results
The deep-learning segmentation took less than 1 second, while manual segmentation took
approximately 10 seconds per image.
The M1’s models consistently achieved equal or better scores (SegNet F1:0.80±0.08) than G2's
manual segmentations (F1:0.68±0.25). G2's low F1/DICE and precision scores indicate poor
agreement between the manual contours. Models from M2 achieved lower scores than G2 and
M1’s models since they demonstrated a strong bias towards predicting no tumor for all images.
Conclusion
Future studies will investigate training on an equal number of images with/without tumor, using
ground truth contours from multiple experts simultaneously.
RapidBrachyMCTPS: An open-source dose calculation and optimization tool for brachytherapy research Presentation
COMP, 01.06.2021.
@misc{Morcos2021c,
title = {RapidBrachyMCTPS: An open-source dose calculation and optimization tool for brachytherapy research},
author = {Marc Morcos and Majd Antaki and Alana Thibodeau-Antonacci and Jonathan Kalinowski and Harry Glickman and Shirin A. Enger},
year = {2021},
date = {2021-06-01},
howpublished = {COMP},
keywords = {},
pubstate = {published},
tppubtype = {presentation}
}
2021.
@award{Thibodeau-Antonacci2021b,
title = {Development of a Dynamic Shielding Intensity-Modulated Brachytherapy Applicator for the Treatment of Rectal Cancer},
author = {Alana Thibodeau-Antonacci and Té Vuong and Hamed Bekerat and Liheng Liang and Shirin A. Enger},
url = {https://curietherapi.es/},
year = {2021},
date = {2021-05-23},
urldate = {2021-05-23},
organization = {Curietherapies},
abstract = {Oral presentation given online at the annual congress of Curietherapies https://curietherapi.es/},
howpublished = {Annual Congress of Curietherapies},
keywords = {},
pubstate = {published},
tppubtype = {award}
}
Oral presentation given online at the annual congress of Curietherapies https://curietherapi.es/
OC-0112 development of a dynamic-shielding intensity modulated endorectal brachytherapy applicator Presentation
Radiotherapy and Oncology, 01.05.2021, ISBN: 0167-8140, 1879-0887.
@misc{Thibodeau-Antonacci2021,
title = {OC-0112 development of a dynamic-shielding intensity modulated endorectal brachytherapy applicator},
author = {Alana Thibodeau-Antonacci and Té Vuong and Hamed Bekerat and Lilian Childress and Shirin A. Enger},
url = {https://www.thegreenjournal.com/article/S0167-8140(21)06316-7/fulltext},
doi = {10.1016/S0167-8140(21)06316-7},
isbn = {0167-8140, 1879-0887},
year = {2021},
date = {2021-05-01},
abstract = {www.thegreenjournal.com},
howpublished = {Radiotherapy and Oncology},
keywords = {},
pubstate = {published},
tppubtype = {presentation}
}
