Click here to

Session: Imaging General ePoster Viewing [Return to Session]

Segmentation of Tumor and Organs at Risk for CBCT-Based Online Adaptive Radiotherapy Using Recurrent Neural Networks with Multi-Scale Memory

H Zhao*, X Liang, B Meng, M Dohopolski, B Choi, B Cai, M Lin, T Bai, D Nguyen, S Jiang, UTSW, Dallas, TX


PO-GePV-I-9 (Sunday, 7/10/2022)   [Eastern Time (GMT-4)]

ePoster Forums

Purpose: Accurate and automated segmentations of targets and organs at risk (OARs) are critical for the successful clinical implementation of online adaptive radiotherapy (oART). Our current workflow transposes the planning CT to online CBCT via a hybrid of AI contouring and deformable registration. However, this workflow demands high CBCT image quality or successful registration, which can be challenging in clinic. In this work, we propose to utilize contours and images from all previous adaptive treatments (per patient) to improve segmentation of the latest fraction.

Methods: We built a Recurrent Neural Networks (RNNs) with Long Short-Term Memory (LSTM) to exploit the prior information (ie prior segmentations) regarding an individual patient. This RNNs is based on a typical U-Net with 3D convolution, with LSTM units inserted to each down-sampled block. The LSTMs within each down-sampled block preserved a multi-scale memory about the current patient, and the memory was updated with each online fraction. The input of RNNs included an image and clinician-edited contours of the most recent fraction, and an image of the current fraction. The label was the clinician-edited contours for the current fraction. The clinician-edited contours of the most recent fraction and the U-Net with the same input was used to benchmark RNNs performance. Contour accuracy was measured using Dice coefficient.

Results: The models were tested on 4 patients from a clinical oART system and each patient has 6 fractions. Structures of left brachial plexus, esophagus, left submandibular gland, left parotid and GTV-nodal were tested. Respecting the structures, Dice coefficients of the proposed RNNs minus input are 3.6%, 3.7%, 4.1%, 2.9%, 0.4% and Dice coefficients of the RNNs minus U-Net are 1.3%, 1.1%, 1.1%, 1.0%, 0.3%.

Conclusion: The proposed RNNs achieved higher accuracy in segmentation of tumor and OARs for online adaptive radiotherapy.

Funding Support, Disclosures, and Conflict of Interest: Two research grants from Varian given to Steve Jiang


Segmentation, Cone-beam CT


IM/TH- Image Segmentation Techniques: Machine Learning

Contact Email