(Sunday, 7/14/2019)
Room: ePoster Forums
Purpose: To verify the adaptability of artificial neural network to various types of tumor delineation using PET-CT based deep learning tumor delineation system.
Methods: The tumor delineation system was developed based on U-net structure which is frequently used in semantic segmentation field. Input data of designed system is adjoined PET-CT image(256×256×2) and output is tumor labeled data. 834 PET-CT slices consisting of 402 Head and neck (HnN) and 432 sarcoma were collected from the cancer image archive (TCIA) to train the developed system. Randomly selected 680 images from the total dataset were used for training and the rest were used for validation. Gross tumor volume was delineated through the clinical data contained in the collected dataset. While the system training, data was augmented in every epoch to avoid overfitting. Training process was performed through the ADAM optimizer and training loss was calculated using root-mean-square value. Other parameters like learning late were changed during several training processes. System accuracy was quantified through the dice similarity coefficient between output and ground truth.
Results: Trained system showed considerable DSC value in sarcoma segmentation. Calculated mean DSC of HnN case was lower than sarcoma case in the test dataset. Not limited to HnN case, System output tended to be more misclassified when the tumor size is smaller.
Conclusion: This work indicates that artificial neural network based tumor delineation system can be extended to multi label classification likewise other semantic segmentation systems. To improve the system accuracy, more training data and novel network structure development will performed in further study.
Keywords
Taxonomy
IM/TH- image segmentation: Multi-modality segmentation
Contact Email
