Click here to

Session: [Return to Session]

A Clinical-Friendly Deep Interactive Segmentation Algorithm for Volumetric Image

T Bai1*, M Lin1, X Liang1, B Wang1,2, M Dohopolski1, B Cai1, D Nguyen1, S Jiang1, (1) The University of Texas Southwestern Medical Ctr, Dallas, TX, (2) Southern Methodist University, Dallas, TX

Presentations

PO-GePV-M-15 (Sunday, 7/10/2022)   [Eastern Time (GMT-4)]

ePoster Forums

Purpose: Interactive medical image segmentation is vital in clinics to improve contour revision efficiency. Deep learning (DL)-based methods have been investigated, achieving state-of-the-art performance. However, most existing techniques are developed for 2D images, which cannot be directly applied to clinically dominant 3D images. This work is to design a clinical-friendly 3D image interactive segmentation algorithm by respecting the clinicians’ revision habits.

Methods: In current clinical practice, the clinicians get used to finishing the contour revision of one whole slice before going to the next one, and those revised slices should not be changed anymore. To respect these two facts, we designed a new DL model with dynamic input and trained it with a dynamic target. Specifically, we feed into the network a four-channel input composing of the image, initial/current mask, revision hints, and dynamically updated revised contours. In the training phase, we use a dynamic target as the supervision signal, where the ground truth contours of those revised slices are replaced with the simulated revised contours. This mechanism lets the network learn an identity mapping for those revised slices and thus won’t change them anymore. Besides, we proposed a loss modulation module such that the gradient can be dominated by the error from the current revising slice and thus guide its correction. We employed the open-accessed StructSeg dataset to validate the proposed method and used the Dice Coefficient (DSC) and clicks per slice for performance quantification.

Results: Experimental results show that the proposed method can improve the DSC from 0.843 (initial) to 0.873, given 2.45 clicks per slice on average. The revised contours are ideally kept unchanged when the model updates the other slices.

Conclusion: We proposed a new DL-based 3D image interactive segmentation algorithm which is more analogous to the clinicians’ contour revision habits.

Keywords

Segmentation

Taxonomy

IM- CT: Machine learning, computer vision

Contact Email

Share: