Automatic 3D Segmentation of Head and Neck Tumours and Cancerous Lymph Nodes


Session type:


Baixiang Zhao1,John Soraghan1,Gaetano Di-caterina1,Derek Grose2
1university of strathclyde,2Beatson West of Scotland Cancer Centre



The implementation of intensity-modulated radiotherapy (IMRT) has allowed improved conformity of high dose radiation to the target volumes. However, there are concerns such as a significant increase in clinician voluming time and complexity which can potentially lead to issues with accuracy of delineation.

 There is emerging interest in the role of automatic delineation tools to improve time, efficiency and variations when compared with manual delineation. Unfortunately to date techniques have not been particularly successful in defining pathological disease having being more successful in the role of defining normal anatomical structures.

As such there is a significant clinical need to develop and validate automatic tools to aid in the delineation of the GTV.


 This work utilised MRI data from Beatson West of Scotland Centre, Glasgow to build a framework for 3D head and neck tumour segmentation. A novel algorithm called level set method extracted the 3D volume of throat, cancerous lymph node (LN) and tumour.  These automatic results were then compared against contouring performed by 2 expert head/neck oncologists.     


The proposed framework was tested on MRI dataset consisted of 17 patients with head and neck cancers. In experiments, the proposed method could fully automatically segment majority 3D volume of targets. And, compared to consensus manual delineation, the automatic method could achieve over 0.85 average Dice similarity coefficients (DSC) on throat segmentation, and about 0.75 on LNs, and 0.70 on tumours.


In conclusion, this work built a fully automatic 3D head and neck tumour segmentation algorithm to improve radiation treatment planning. The experiments on real medical data shows that this algorithm provides acceptable results and less time consumption.