(Tuesday, 7/16/2019) 10:00 AM - 10:30 AM
Room: Exhibit Hall | Forum 6
Purpose: Automatic and accurate organ localization can be an important issue for medical image analysis and radiation therapy. Among the abdominal organs, pancreas localization/segmentation can be difficult since this small, soft and flexible organ demonstrates high inter-patient variability in terms of shape and volume. More recently, deep learning based techniques have achieved plausible accuracy in organ localization/segmentation problems. In this work, we conduct pancreas localization using convolution neural network on the open source data-set from NIH.
Methods: Pancreas localization is achieved by estimating the coordinates of a bounding box. The proposed method includes two steps. Firstly, U-Net architecture was employed for the supervised training of pancreas segmentation. The loss function is composed of two terms: dice coefficient loss and center point loss. The center point is denoted as the weighted sum of the coordinates of labels. In the second step, the width and height of the bounding box can be extracted from the estimated label maps in the first step. Given the estimated center point coordinates, the width and the height of the bounding box, it is straightforward to calculate the bounding box coordinates.
Results: The model is trained and evaluated on the pancreas-CT data-set from NIH clinical center, which includes 82 abdominal contrast enhanced CT scans from 53 male and 27 female subjects. The IoU of the bounding box estimation achieve 0.87 on the data-set. The dice coefficient for the segmentation is 0.66 on the data-set.
Conclusion: Due to the close relationship between localization and segmentation, we are able to achieve a good accuracy of pancreas localization with mild segmentation accuracy. Moreover, pancreas localization can be super-fast, which makes it easily and promising to be integrated into clinical workflow. Future works include to improve the segmentation accuracy of pancreas with deep learning techniques.
Funding Support, Disclosures, and Conflict of Interest: NIH R44CA183390 NIH R01CA188300
Keywords
Not Applicable / None Entered.
Taxonomy
Not Applicable / None Entered.
Contact Email
