MENU

Click here to

×

Are you sure ?

Yes, do it No, cancel
  1. Home
  2. Meeting Program
  3. All Sessions
  4. All Multi-Disciplinary General ePoster Discussion
  5. Clinical Evaluation of Atlas and Deep Learning Based Automatic Contouring for Nasopharyngeal Carcinoma

Clinical Evaluation of Atlas and Deep Learning Based Automatic Contouring for Nasopharyngeal Carcinoma

J Wang1 , C Yang2 , B Qu1 , L Ma1 , W Fan1 , B Liu3 , S Xu1,4*, (1)Department of Radiation Oncology,PLA General Hospital ,Beijing,100853,China (2) Manteia Medical Technologies, Milwaukee, WI 53226. (3)Image Processing Center,Beihang Unoversity, Beijing, 100191,China (4) Beijing Advanced Innovation Center for Big Data-Based Precision Medicine, Beihang University, Beijing, 100083, China

Presentations

(Wednesday, 7/17/2019) 10:30 AM - 11:00 AM

Room: Exhibit Hall | Forum 5

Purpose: To investigate the difference between atlas-based and deep learning (DL) based segmentation of organs at risk (OARs) for nasopharyngeal carcinoma.

Methods: One set of atlas consisting of 120 cases of nasopharyngeal carcinoma were established in MIM-Maestro (Ver 6.6.5). 14 OARs on 20 patients of nasopharyngeal carcinoma were contoured by experience physicians based on published consensus guidelines and were defined as reference volumes, Vref. The 20 patients were selected randomly outside of the atlas sets. These OARs were also auto-contoured using atlas-based model (in MIM) and deep learning-based model ( in AccuContour®, Manteia Medical Technologies ), and the volumes were defined collectively as Vatlas and VDL. The similarities between Vatlas�VDL and Vref were assessed using dice similarity coefficient (DSC), Jaccard coefficient(JC), maximum Hausdorff distance(HDmax) and deviation of centroid. A paired t-test was performed to show the significance of the difference.

Results: For the brainstem, eyeballs, and lens, the results of DL contouring were significantly better than those of Vatlas (p<0.05). For inner ears, temporo-mandibular joints and the oral cavity, Vatlas seemed better (p<0.05). For optic nerves, there were no significant difference between the two methods (p>0.05). Spatial volumes of VDL of parotid glands were closer than Vatlas to those of Vref.

Conclusion: Deep learning based segmentation performs significantly better than atlas-based segmentation on the contouring accuracy of OARs when clear boundaries of those OARs are present, while for those OARs with unclear boundaries, the atlas-based method performs better. A possible reason could be that the two models were built from two different groups of labelled data while they were labelled with different software tools and different groups of people, and the same group of physicians labelled Vatlas and Vref. DL based method catches well the edge information when data labelling is less subjective to human.

Funding Support, Disclosures, and Conflict of Interest: This work was supported in part by the grant from National Key R&D Program of China (No. 2017YFC0112100) and National Natural Science Foundation of China(No.61601012&81801799)

Keywords

Not Applicable / None Entered.

Taxonomy

Not Applicable / None Entered.

Contact Email

Additional Links