Click here to

Session: AI/ML Autoplanning, Autosegmentation, and Image Processing II [Return to Session]

Automatic Image and Contour Augmentation for Deep Learning Auto-Segmentation of Complex Anatomy

N Dang*, Y Zhang, A Amjad, J Ding, C Sarosiek, X Li, Medical College of Wisconsin, Milwaukee, WI

Presentations

SU-F-BRB-3 (Sunday, 7/10/2022) 2:00 PM - 3:00 PM [Eastern Time (GMT-4)]

Ballroom B

Purpose: Deep learning auto-segmentation (DLAS) has limited success for complex anatomy, such as organs in abdomen, thus, requires larger datasets for its improvement. This work aims to develop an automated data augmentation tool (ADAT) to substantially enlarge the variation and size of image and contour datasets to improve DLAS for abdominal organs on MRI.

Methods: The ADAT was developed to transform existing dataset of images and associated contours in two steps: (1) geometry augmentation: including 3D affine transformations (shifting, rotation, and scaling), and elastic B-spline deformation, and (2) image quality augmentation: random values drawn from a Gaussian distribution (noise) or a 3rd order polynomial (bias field). The tool accepts DICOM input of both images and contours, creates different mask arrays with the same size as the images from each organ contours, and simultaneously augments images and contours by a user defined augmentation factor and writes out the results in the same format as the input. To test ADAT, a dataset of 65 MRIs and contours of both ground truth and DLAS contours was augmented. The augmented datasets were used to train DLAS of abdominal organs and a deep learning automatic contour refinement model.

Results: The ADAT successfully created 650 augmented datasets from the 65 input datasets with an augmentation factor of 10. There is no statistical difference between the augmented DLAS contours and the DLAS contours directly from the augmented images. Performance of the DLAS model trained with the 650 augmented datasets was improved comparing to the DLAS model trained with the original 65 datasets, with Dice similarity coefficient increased by 5% for bowels and stomach, and 20% for duodenum and pancreas.

Conclusion: The newly developed automatic data augmentation tool can successfully enlarge image and contour datasets for DLAS or other deep learning approaches, improving auto-segmentation for complex anatomy.

Keywords

Segmentation, Image Processing, MRI

Taxonomy

IM/TH- image Segmentation: General (Most aspects)

Contact Email

Share: