Room: AAPM ePoster Library
Purpose: guidance has been widely used in radiation therapy. Correctly identifying anatomical landmarks, like T12 vertebra, is the key to success. Until recently, the detection of those landmarks still requires tedious manual inspections and annotations; and superior-inferior misalignment to the wrong vertebral body is still a relatively common IGRT error. It is necessary to develop an automated approach to detect those landmarks on CT and CBCT images.
Methods: designed a novel network to detect anatomic structures from tomography data, requiring only a small amount of training data. The model is trained from thirty-five abdominal CT images scanned with the same imaging protocol. Then the model was tested on fourteen CT images and six CBCT images. The test CT images were from two scanning protocols (eleven liver patients and three lung patients). No test images were used during training.
Results: used the distance of T12 centers between prediction and ground-truth as detection error for IGRT safety check. On all fourteen test CT images, T12s are accurately identified with a mean detection error of 3.97 mm. In the superior-inferior direction, the mean absolute detection error is 0.20 mm. On six CBCT images, only one T12 in a scan was identified as non-existence. All the others were identified correctly with a mean detection error of 4.49 mm. It took about 3.30 seconds to run the detection on a image dataset of 416?288?128 voxels.
Conclusion: our study, our proposed approach demonstrates the capability of accurately detecting structures with high similarity from a small amount of annotated training data. Furthermore, the model is robust to work with the high noise level of CBCT. It has great potential to be integrated into IGRT workflow to improve safety and minimize gross alignment errors.
Not Applicable / None Entered.