Deep learning for real-time multi-label segmentation 3D Ultrasound volumes of tongue tumor specimens

Scientific Research Project Number: MA 2021 08
Place: Netherlands Cancer Institute, Antoni van Leeuwenhoek Hospital, Department of Head and Neck Surgery and Oncology, Verwelius 3D Lab

Introduction

Although an oncological surgeon aims to remove tongue cancer with a margin of no less than 5 mm, 85% of margins are reported as close or positive (<5mm). Ultrasound (US) proves to be a feasible technique to assess resection margins intra-operatively (1). By reconstructing the 2D US images into a 3D US volume, it is able to assess the entire specimen. However, segmentation of all single slices in the 3D US volume is time-consuming and therefore not feasible at the operating room. By applying 3D volume segmentation of healthy tissue and tumour enables fast and accurate intra-operative as-sessment of the entire resected tongue specimen. Several studies show the possibilities of segmen-tation of US by deep learning: one segmenting the tongue contour in 2D US images(2), and another segmenting the placenta in 3D US volumes (3,4).


Description of the SRP Project/Problem

This study aims to create a fast and accurate segmentation model for 3D US volumes. The model requires intra-operative usability, and preferably combined/integrated in visualization software.


Research questions

- Can a trained multi-label segmentation model of 3D US volumes reach accuracy rates sufficient for clinical setting?

- Who can a trained model be improved and trained as data acquisition continues in this prospective study?

- What is required to implement a trained segmentation model in clinical setting?


Expected results

A trained segmentation model able to automatically segment 3D US volumes with clinical accuracy.

Implementation of the trained segmentation model for clinical use at the operating room.

Tools to improve the model during the prospective study.


Time period:

7 months


Contact:

Nicolaas Bekedam, NKI-AVL, department of Head and Neck Surgery and Oncology, n.bekedam@nki.nl


References

1. Brouwer de Koning SG, Karakullukcu MB, Lange CAH, Schreuder WH, Karssemakers LHE, Ruers TJM. Ultrasound aids in intraoperative assessment of deep resection margins of squamous cell carcinoma of the tongue. Br J Oral Maxillofac Surg [Internet]. 2020;58(3):285–90. Available from: https://doi. org/10.1016/j.bjoms.2019.11.013

2. Zhu J, Styler W, Calloway I. A CNN-based tool for automatic tongue contour tracking in ultrasound images. arXiv. 2019;1–6.

3. Looney P, Stevenson GN, Nicolaides KH, Plasencia W, Molloholli M, Natsis S, et al. Automatic 3D ultrasound segmentation of the first trimester placenta using deep learning. Proc - Int Symp Biomed Imaging. 2017;279–82.

4. Looney P, Stevenson GN, Nicolaides KH, Plasencia W, Molloholli M, Natsis S, et al. Fully automated, real-time 3D ultrasound segmentation to estimate first trimester placental volume using deep learning. JCI insight. 2018;3(11):1–9.


Contact

Nicolaas Bekedam, NKI-AVL, department of Head and Neck Surgery and Oncology, n.bekedam@nki.nl