MENU

Click here to

×

Are you sure ?

Yes, do it No, cancel

A Generalizable Contour Validation Method Using Deep Learning-Based Image Classification

Y Zhang*, F Ceballos, Y Liang, L Buchanan, X Li, Medical College of Wisconsin, Milwaukee, WI

Presentations

(Thursday, 7/16/2020) 2:00 PM - 3:00 PM [Eastern Time (GMT-4)]

Room: Track 5

Purpose:
Manual validation of auto-segmented contours can be labor-intensive and time-consuming. Previously reported automatic contour quality assurance(ACQA) techniques may work for only limited organs and/or image modality. We propose a deep learning based ACQA method that is generalizable and implementable.

Methods:
Our method includes two major stages: 1)Image pre-processing consisting of normalizing, cropping and masking, and generating three-channel RGB images by combining the target slice with the top and bottom neighboring slices; 2)augmenting and feeding the processed images into a convolutional neural network(CNN) image classification model for training. A dataset of T2-weighted MRIs of 41 head&neck cancer patients along with contours of the parotid glands and submandibular glands was utilized (28 for training, 5 for validation, and 8 for testing). Each MRI included three sets of contours, one accurate set by manual delineation and two inaccurate sets created manually and automatically by deformable image registration(DIR). A slice was labeled as inaccurate if the mean distance to agreement=2 mm or Hausdorff distance=10 mm compared with the corresponding accurate contour. The labels were then manually checked to avoid mislabeling. The model was tested with the three sets contours. The sensitivity, specificity, and accuracy were used to measure the models’ performance.

Results:
The average sensitivity, specificity and accuracy to identify inaccurate or accurate contours were 97±2.2%, 89±5.1% and 94±1.3% for the manually created contours and 83±4.5%, 84±5.9% and 84±1.7% for the DIR auto-generated contours. Using a workstation equipped with a i7-6700 CPU, the model training took 4 hours for each organ, but contour QA took less than 1 second for a test case.

Conclusion:
The proposed DL-ACQA method can quickly and automatically identify accurate and inaccurate contour slices, eliminating the labor-intensive manual review process. This technique based on CNN can be applied to any organ thus may be implemented for routine clinical use.

Keywords

Quality Assurance, Segmentation, Decision Theory

Taxonomy

IM/TH- Formal Quality Management Tools: Machine Learning

Contact Email