Click here to


Are you sure ?

Yes, do it No, cancel

Weakly-Supervised Deep Learning Based Automatic Image Segmentation Via Deformable Image Registration

W Chi123*, W Lu1, L Ma1, J Wu1, H Chen4,, M Tan23, X Gu1, (1) UT Southwestern Medical Center, Dallas, TX, (2) South China University Of Technology, Guangzhou, China, (3) Guangzhou Laboratory, Guangzhou, China, (4) Sun Yat-sen University, Guangzhou, China


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

Room: AAPM ePoster Library

Purpose: Deep learning (DL) based auto-segmentation requires amounts of clean labeled data to train a robust model. However, in medical field, clean labeled data are limited, while images without annotation are abundant. To alleviate the influence of limited labels issue, we propose to develop a weakly-supervised deep learning model via deformable image registration (DIR) labeled-data generation.

Methods: Our approach is based on two assumptions 1) a handful of atlases (images with ground-truth contours) and numerous non-contoured images are available; 2) DL models are robust to noise. The entire approach consists of two steps: 1) DIR to propagate contours from the atlas to unlabeled images to generate pseudo-contours. 2) A 3D recursive ensemble organ segmentation (REOS) model is supervised trained with generated pseudo-contours.
The developed approach is evaluated on 31 head-and-neck CT images and contours set from TCIA database. The entire dataset is split into three sets: training (20), validation (4), and testing (7). For each training image, 19 pseudo-contour sets are deformed from the rest one as the atlas. Totally 380 generated pseudo-contour sets are used to train REOS model. The trained DL model is evaluated using Dice coefficient (DC) on 7 testing cases’ 5 OARs, including mandible, L&R parotids and L&R submandibular glands. DL segmentation is compared with DIR-average and DIR-majority-voting, which are the mean or majority voting DC of 140 DIR generated contours, respectively.

Results: Our model achieves DCs of 86.2%±1.8%, 75.7%±3.9%, 72.2%±3.7%, 64.4%±6.1% and 59.9%±4.7% on mandible, L&R parotids and L&R submandibular glands, respectively. Our model outperforms DIR-average and DIR-majority-voting by 12% and 2.7%, respectively.

Conclusion: We developed a weakly-supervised DL algorithm for segmentation on datasets with very limited labels. Results demonstrate the weakly-supervised DL method outperforms traditional multi-atlas DIR methods and is promising for DL-based limited data medical image segmentation application.

Funding Support, Disclosures, and Conflict of Interest: This study was supported by a NIH grant (R01 CA235723).


Computer Vision, Data Acquisition, Image Processing


IM/TH- image Segmentation: CT

Contact Email