(Sunday, 7/12/2020) [Eastern Time (GMT-4)]
Room: AAPM ePoster Library
Purpose:
To compare the performance of deep-learning (DL)-based and atlas-based autosegmentation in a clinical cohort of breast cancer patients.
Methods:
In this retrospective study, a clinical cohort of 400 breast cancer patients was included. Based on CTs and clinical delineations, DL-models were trained using the Admire 3.7 software (Elekta). Data were split 80/20 for model training and evaluation.
A multi-structure model was trained to segment heart, patient-outline, left and right lung. For structures with missing delineations, separate single-structure models were trained and evaluated. This was done for the spinal cord, left and right humeral head, left breast and right breast, using respectively 177, 100, 91, 32 and 23 patients.
The performance of DL-based autosegmentation was compared to that of our clinically used atlas-based autosegmentation (Admire 2.5, Elekta). To this end, contour overlap (Dice-coefficient) and maximal distance (99?? percentile of Hausdorff-distance) was calculated between both autosegmentation methods and the clinical delineation.
Results:
Dice-coefficients were comparable; on average 0.94 for both methods. Improvements of using DL-based compared to atlas-based segmentations were seen in the extent of local miss-segmentation. On average over all patients, DL-based autosegmentation reduced maximal distances from; 7.4 to 6.2 mm for heart, 13.2 to 4.4 mm for left lung, 12.6 to 5.3 mm for right lung, 12.6 to 8.5 mm for spinal cord, 6.1 to 4.9 mm for right humeral head and 16.6 to 13.3 mm for left breast. For the left humeral head, right breast and patient-outline, the average maximal distance increased from; 2.9 to 5.3, 23.0 to 23.3 and 3.1 to 3.4 mm, respectively.
Conclusion:
DL-based autosegmentation outperformed the current clinically used atlas-based autosegmentation for most structures. DL-based autosegmentation reduced the extent of local miss-segmentation. We expect that this leads to less manual editing of autosegmentations in clinical practice.
Funding Support, Disclosures, and Conflict of Interest: The Erasmus MC has a research agreement with Elekta AB.
Keywords
Not Applicable / None Entered.
Taxonomy
IM/TH- Image Segmentation Techniques: Modality: CT
Contact Email