Learned Delineation of Gross Tumor Volume Incorporating Intra-Observer Variability
AAPM ePoster Library. Marin T. 07/12/20; 302433; PO-GeP-M-280
Topic: Science Council Session - Data-Driven Automation and Decision Making
Dr. Thibault Marin
Contributions
Contributions
Abstract
Poster Number: PO-GeP-M-280
Abstract ID: 52720
Learned Delineation of Gross Tumor Volume Incorporating Intra-Observer Variability
T Marin*, C Ma, R Lahoud, F Xing, P Wohlfahrt, M Moteabbed, J Woo, X Ma, K Grogg, Y Chen, G El Fakhri, Massachusetts General Hospital, Boston, MA
Multi-Disciplinary General ePoster
Category: Scientific:Multi-Disciplinary:Science Council Session - Data-Driven Automation and Decision Making:Data-Driven Decision Making
Purpose: To develop a deep learning-based method for delineation of the gross tumor volume (GTV) in computed tomography (CT) scans of patients with soft tissue sarcomas.
Methods: CT scans were acquired for 15 patients with soft tissue sarcomas and combined with publicly available datasets from The Cancer Imaging Archive (McGill University, Canada) for a total of 45 patients. GTV was specified three times by a clinician in three independent sessions. The intersection (consensus), exclusive disjunction (partial agreement) and non-disjunction (healthy tissue) of the three GTV contours were used to classify CT voxels. A deep convolutional neural network based on the 2D U-Net architecture with attention filters in the expanding path was designed to learn the classification and thus the intra-observer variability in GTV delineation. The network was trained to minimize an ordinal cross-entropy loss. Five datasets were excluded from the training and used only for validation.
Results: Segmentation metrics were computed for both the consensus and the partial agreement contours. Results showed good agreement with the reference contours. The median Dice score, sensitivity and specificity for the consensus region were 0.77, 0.73 and 0.99 respectively. For the partial agreement region, the median metrics were 0.77, 0.82, 0.99. The Dice scores ranged from 0.48 to 0.91 (partial agreement contour), which approaches the measured intra-reader variability (median Dice score: 0.92). The median of the average Hausdorff distance was 3.16 mm and 5.72 mm for the consensus and partial agreement contours respectively.
Conclusion: The proposed deep convolutional neural network is able to accurately and reliably localize the GTV and integrate its intra-observer variability in contrast to other automatic segmentation approaches using only a single contour. This can be a valuable tool to clinicians in GTV delineation by accurately identifying highly confident as well as uncertain tumor regions.
Taxonomy:IM/TH- Image Segmentation Techniques: Machine Learning
Keywords: segmentation,treatment planning,image analysis,
Abstract ID: 52720
Learned Delineation of Gross Tumor Volume Incorporating Intra-Observer Variability
T Marin*, C Ma, R Lahoud, F Xing, P Wohlfahrt, M Moteabbed, J Woo, X Ma, K Grogg, Y Chen, G El Fakhri, Massachusetts General Hospital, Boston, MA
Multi-Disciplinary General ePoster
Category: Scientific:Multi-Disciplinary:Science Council Session - Data-Driven Automation and Decision Making:Data-Driven Decision Making
Purpose: To develop a deep learning-based method for delineation of the gross tumor volume (GTV) in computed tomography (CT) scans of patients with soft tissue sarcomas.
Methods: CT scans were acquired for 15 patients with soft tissue sarcomas and combined with publicly available datasets from The Cancer Imaging Archive (McGill University, Canada) for a total of 45 patients. GTV was specified three times by a clinician in three independent sessions. The intersection (consensus), exclusive disjunction (partial agreement) and non-disjunction (healthy tissue) of the three GTV contours were used to classify CT voxels. A deep convolutional neural network based on the 2D U-Net architecture with attention filters in the expanding path was designed to learn the classification and thus the intra-observer variability in GTV delineation. The network was trained to minimize an ordinal cross-entropy loss. Five datasets were excluded from the training and used only for validation.
Results: Segmentation metrics were computed for both the consensus and the partial agreement contours. Results showed good agreement with the reference contours. The median Dice score, sensitivity and specificity for the consensus region were 0.77, 0.73 and 0.99 respectively. For the partial agreement region, the median metrics were 0.77, 0.82, 0.99. The Dice scores ranged from 0.48 to 0.91 (partial agreement contour), which approaches the measured intra-reader variability (median Dice score: 0.92). The median of the average Hausdorff distance was 3.16 mm and 5.72 mm for the consensus and partial agreement contours respectively.
Conclusion: The proposed deep convolutional neural network is able to accurately and reliably localize the GTV and integrate its intra-observer variability in contrast to other automatic segmentation approaches using only a single contour. This can be a valuable tool to clinicians in GTV delineation by accurately identifying highly confident as well as uncertain tumor regions.
Taxonomy:IM/TH- Image Segmentation Techniques: Machine Learning
Keywords: segmentation,treatment planning,image analysis,
Poster Number: PO-GeP-M-280
Abstract ID: 52720
Learned Delineation of Gross Tumor Volume Incorporating Intra-Observer Variability
T Marin*, C Ma, R Lahoud, F Xing, P Wohlfahrt, M Moteabbed, J Woo, X Ma, K Grogg, Y Chen, G El Fakhri, Massachusetts General Hospital, Boston, MA
Multi-Disciplinary General ePoster
Category: Scientific:Multi-Disciplinary:Science Council Session - Data-Driven Automation and Decision Making:Data-Driven Decision Making
Purpose: To develop a deep learning-based method for delineation of the gross tumor volume (GTV) in computed tomography (CT) scans of patients with soft tissue sarcomas.
Methods: CT scans were acquired for 15 patients with soft tissue sarcomas and combined with publicly available datasets from The Cancer Imaging Archive (McGill University, Canada) for a total of 45 patients. GTV was specified three times by a clinician in three independent sessions. The intersection (consensus), exclusive disjunction (partial agreement) and non-disjunction (healthy tissue) of the three GTV contours were used to classify CT voxels. A deep convolutional neural network based on the 2D U-Net architecture with attention filters in the expanding path was designed to learn the classification and thus the intra-observer variability in GTV delineation. The network was trained to minimize an ordinal cross-entropy loss. Five datasets were excluded from the training and used only for validation.
Results: Segmentation metrics were computed for both the consensus and the partial agreement contours. Results showed good agreement with the reference contours. The median Dice score, sensitivity and specificity for the consensus region were 0.77, 0.73 and 0.99 respectively. For the partial agreement region, the median metrics were 0.77, 0.82, 0.99. The Dice scores ranged from 0.48 to 0.91 (partial agreement contour), which approaches the measured intra-reader variability (median Dice score: 0.92). The median of the average Hausdorff distance was 3.16 mm and 5.72 mm for the consensus and partial agreement contours respectively.
Conclusion: The proposed deep convolutional neural network is able to accurately and reliably localize the GTV and integrate its intra-observer variability in contrast to other automatic segmentation approaches using only a single contour. This can be a valuable tool to clinicians in GTV delineation by accurately identifying highly confident as well as uncertain tumor regions.
Taxonomy:IM/TH- Image Segmentation Techniques: Machine Learning
Keywords: segmentation,treatment planning,image analysis,
Abstract ID: 52720
Learned Delineation of Gross Tumor Volume Incorporating Intra-Observer Variability
T Marin*, C Ma, R Lahoud, F Xing, P Wohlfahrt, M Moteabbed, J Woo, X Ma, K Grogg, Y Chen, G El Fakhri, Massachusetts General Hospital, Boston, MA
Multi-Disciplinary General ePoster
Category: Scientific:Multi-Disciplinary:Science Council Session - Data-Driven Automation and Decision Making:Data-Driven Decision Making
Purpose: To develop a deep learning-based method for delineation of the gross tumor volume (GTV) in computed tomography (CT) scans of patients with soft tissue sarcomas.
Methods: CT scans were acquired for 15 patients with soft tissue sarcomas and combined with publicly available datasets from The Cancer Imaging Archive (McGill University, Canada) for a total of 45 patients. GTV was specified three times by a clinician in three independent sessions. The intersection (consensus), exclusive disjunction (partial agreement) and non-disjunction (healthy tissue) of the three GTV contours were used to classify CT voxels. A deep convolutional neural network based on the 2D U-Net architecture with attention filters in the expanding path was designed to learn the classification and thus the intra-observer variability in GTV delineation. The network was trained to minimize an ordinal cross-entropy loss. Five datasets were excluded from the training and used only for validation.
Results: Segmentation metrics were computed for both the consensus and the partial agreement contours. Results showed good agreement with the reference contours. The median Dice score, sensitivity and specificity for the consensus region were 0.77, 0.73 and 0.99 respectively. For the partial agreement region, the median metrics were 0.77, 0.82, 0.99. The Dice scores ranged from 0.48 to 0.91 (partial agreement contour), which approaches the measured intra-reader variability (median Dice score: 0.92). The median of the average Hausdorff distance was 3.16 mm and 5.72 mm for the consensus and partial agreement contours respectively.
Conclusion: The proposed deep convolutional neural network is able to accurately and reliably localize the GTV and integrate its intra-observer variability in contrast to other automatic segmentation approaches using only a single contour. This can be a valuable tool to clinicians in GTV delineation by accurately identifying highly confident as well as uncertain tumor regions.
Taxonomy:IM/TH- Image Segmentation Techniques: Machine Learning
Keywords: segmentation,treatment planning,image analysis,
{{ help_message }}
{{filter}}