MENU

Click here to

×

Are you sure ?

Yes, do it No, cancel

Multiple Resolution Residual Network for Automatic Thoracic Organs-At-Risk Segmentation in CT

H Um*, J Jiang, M Thor, A Rimner, L Luo, J Deasy, H Veeraraghavan, Memorial Sloan Kettering Cancer Center, New York, NY

Presentations

(Sunday, 7/12/2020)   [Eastern Time (GMT-4)]

Room: AAPM ePoster Library

Purpose: To implement and evaluate a multiple resolution residual network (MRRN) for multiple normal organs-at-risk (OAR) segmentation from computed tomography (CT) images for thoracic radiotherapy treatment (RT) planning.


Methods: Our approach simultaneously combines feature streams extracted from multiple image resolution and feature levels through residual connections to generate the segmentation. The feature streams at each level are updated as the images are passed through various feature levels. The network was trained on 275 thoracic CT scans with 25 CTs held out for validation to segment the heart, left and right lungs, esophagus, and spinal cord. The training dataset was accumulated both internally (planning CTs of 239 patients with locally-advanced non-small cell lung cancer [LA-NSCLC]) and externally (36 CTs provided for training during the 2017 AAPM Thoracic Auto-Segmentation Challenge). All CTs were annotated with expert delineations of the five structures. Training and validation were performed with a total of 25,756 images and 2,645 images, respectively, of size 256x256. The model producing the best average accuracy across all five structures on the validation set was used for testing on the 12 CTs provided during the online testing phase of the AAPM grand challenge. The segmentation accuracy compared to expert delineations was measured using the Dice Similarity Coefficient (DSC) and the performance of our approach was benchmarked against the 3 best-performing methods in the grand challenge.


Results: Median DSC achieved with the MRRN was 0.93 (interquartile range [IQR]: 0.92-0.94) for the heart, 0.97 (IQR: 0.96-0.97) for the left and right lungs, 0.76 (IQR: 0.73-0.80) for the esophagus, and 0.89 (IQR: 0.87-0.90) for the spinal cord.


Conclusion: Our approach for segmenting multiple normal organ structures from CT images outperformed the best-performing method in the AAPM grand challenge for hard-to-segment structures like the esophagus and achieved comparable results for all other structures.

Funding Support, Disclosures, and Conflict of Interest: No related disclosures. Relevant financial activities outside the submitted work: Research grants from: Varian Medical Systems Boehringer Ingelheim Pfizer AstraZeneca Personal fees from: Astrazeneca Merck Research to Practice Cybrexa Non-Financial Support from: Philips/Elekta No stock, patent or other disclosures.

Keywords

CT, Segmentation

Taxonomy

IM/TH- image Segmentation: CT

Contact Email