Y Chen*, S Butler, L Yu, Y Zhou, L Shen, N Kovalchuk, H Bagshaw, M Gensheimer, L Xing, B Han, Stanford University School of Medicine, Stanford, CA

• 
  Y Chen

• 
  
• 
  
• 
  
• 
  
• 
  
• 
  
• 
  

PO-GePV-M-322 (Sunday, 7/10/2022)   [Eastern Time (GMT-4)]

ePoster Forums

Purpose: This study explored patient-specific auto-segmentation via deep learning to facilitate adaptive radiotherapy, based on data of the first two patient cases treated with newly developed RefleXion biology-guided radiotherapy (BgRT) system.

Methods: One head and neck (HaN) case with 17 daily fan-beam CT scans and one prostate case with 26 scans were utilized for patient-specific auto-segmentation. For each body site, a population network was first trained on a population dataset, which contained 67 and 56 patient cases for HaN and pelvis site respectively. Then the pre-trained population network was adapted to the specific RefleXion patient with three different transfer learning strategies, by leveraging the image and contour similarity between the treatment fractions. A longitudinal study was conducted to explore the relationship between auto-segmentation performance and the number of sequential prior data used for learning. The performance of the patient-specific network was compared with the population network and the clinical rigid registration. The corresponding dosimetric impacts resulting from different segmentation and registration methods were also investigated.

Results: The proposed patient-specific network could achieve mean Dice similarity coefficient (DSC) results of 0.875 and 0.905 for HaN and pelvis site respectively, outperforming the population network (0.699 and 0.627) and the registration method (0.724 and 0.716). The DSC of the patient-specific network gradually increased with the increment of longitudinal training cases and approached saturation with more than six training cases. The patient-specific auto-segmentation resulted in more accurate organ mean doses than the registration method, and the dose underestimation in some organs was markedly improved.

Conclusion: Auto-segmentation of RefleXion daily fan-beam CT images was studied based on patient-specific learning for future adaptive radiotherapy. Compared with the common population network and the clinical registration method, the patient-specific network could achieve a much better contouring and dosimetric accuracy and thus is promising to facilitate more precise treatment.

Keywords

Segmentation, Dose, Image-guided Therapy

Taxonomy

IM/TH- Image Segmentation Techniques: Machine Learning

Contact Email