J. Wei1, J. Deng2, and M. Chao3; 1City College of New York, New York, NY, 2Yale University, New Haven, CT, 3Icahn School of Medicine at Mount Sinai, Department of Radiation Oncology, New York, NY
Purpose/Objective(s): Accurate delineation of organs-at-risk (OARs) is time-consuming yet critical in radiotherapy planning of head and neck cancer (HNC). While convolutional neural networks (CNN) such as VGG16/19 have been widely employed as the backbones in the segmentation U-Nets for auto-segmentation of OARs in cancer radiotherapy, a systematic investigation of the newly available Segment Anything Model (SAM) from Meta has been lacking. The goal of this study is to investigate a SAM-based strategy for auto-segmentation of OARs to facilitate radiotherapy treatment planning of HNC patients. Materials/
Methods: CT sets of 42 HNC patients from the HaN-Seg challenge were used in the model training and testing using random 70/30 data splitting. SAM was employed as the deep net to be fine-tuned using the training data. Because the training data for some organs only have <20 slices, randomized data augmentation techniques were used to increase the diversity of the training objects to boost the generalization power of the resulting SAM-based net. The fine-tuning process by 50 epochs was performed and a post-processing step was conducted over the OAR masks generated by the SAM-based model using mathematical morphological operators, which include opening and closing by a circular structural element all over the images, and size filtering based on the component labeling procedure to finalize the OAR segmentation. Twenty-eight OARs were used to fine-tune the model. The contouring accuracy of the proposed SAM model was compared with that of VGG16/19 in terms of recall, precision, and F1 for all the OARs. Results: The performances of SAM are overwhelmingly superior. The averaged precision, recall, and F1 were 0.648 ± 0.052, 0.983 ± 0.031, and 0.778 ± 0.041 (mean ± standard deviation), respectively, for all 28 OARs among twelve patients from the test data. The average time for auto-contouring all 28 OARs simultaneously for each patient was about 4.012 seconds on a Windows 11 PC with an Intel i9-12900KF CPU, an Nvidia RTX 4090 GPU, and 64 GB RAM. With 50 epochs, the SAM model achieved F1 of 0.7-0.9 for all the OARs consistently, regardless of convex or concave shapes. In contrast, the averaged precision, recall, and F1 scores were 0.081 ± 0.13, 0.584 ± 0.392, and 0.121 ± 0.166, respectively for VGG16/19. For convex and compact structures, VGG16/19 can achieve an F1 score of 0.7. However, for concave/elongated/non-connected structures, VGG16/19 failed to generate acceptable segmentations with F1 <0.1 even after over 1000 epochs. Conclusion: Our preliminary results have demonstrated that the developed SAM-based auto-segmentation model can be used to accurately delineate the OARs for efficient radiotherapy planning of HNC patients. More validation will be carried out in our future study.
