Click here to


Are you sure ?

Yes, do it No, cancel

Combining Images and Clinical Diagnostic Information to Improve Automatic Segmentation of Nasopharyngeal Carcinoma Tumors On MR Images

M Cai*, Q Yang, Y Guo, Z Zhang, J Wang, W Hu, C Hu; Department of Radiation Oncology, Fudan University Shanghai Cancer Center, China, Department of Oncology, Shanghai Medical College, Fudan University, China.


(Sunday, 7/12/2020)   [Eastern Time (GMT-4)]

Room: AAPM ePoster Library

Purpose: Identifying tumor margin in magnetic resonance (MR) images for nasopharyngeal carcinoma (NPC) may be difficult. Therefore, it is essential for radiation oncology to accurately and reproducibly delineate tumors from uninvolved tissue. Meanwhile, the information of tumor invasion may be already contained in clinical diagnosis such as T-stage. To improve the performance of tumor segmentation, we incorporate both clinical diagnostic and images information to a new deep learning neural network architecture.

Methods: There were 251 NPC patients enrolled in this study. We extracted patients T-stage from EMR, which was reviewed by one radiation oncologist. The diagnosis MR images including T1-weighted, T2-weighted and contrast-enhanced T1-weighted were collected from PACS. A modified U-net with attention gate was our base network architecture and an additional channel (T-channel) was provided to the network. We randomly divided the whole dataset into training cohort (233 patients) and validation cohort (18 patients). Pytorch was utilized to implement the network and a 10 fold cross validation was performed to get reliable results. Each fold was trained for 600 epochs by two NVIDIA Geforce GTX 1080 Ti GPUs. An Adam optimizer was applied and the learning rate, which was initially set to 10?4, was reduced to 10?5 after 400 epochs. The results of base network and T-stage network were evaluated by the dice similarity coefficient (DSC) and the Jaccard similarity coefficient (JSC).

Results: The average DSC, JSC and their standard deviation (SD) values for the network without T-stage information are 0.804 ± 0.015 (DSC ± SD), 0.691 ± 0.026 (JSC ± SD). For network with T-stage information, the correspond values are 0.833 ± 0.013 (DSC ± SD), 0.719 ± 0.018 (JSC ± SD).

Conclusion: Our proposed T-stage network performs better compare to network use image information only. The unique T-channel effectively utilizes T-stage information to improve the result.


Not Applicable / None Entered.


Not Applicable / None Entered.

Contact Email