MENU

Click here to

×

Are you sure ?

Yes, do it No, cancel

A Patient-Specific Deep-Self-Supervised High-Resolution Method for CT Imaging

Y Lei, Y Fu, T Wang, W Curran, T Liu, X Yang*, Emory Univ, Atlanta, GA

Presentations

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

Room: AAPM ePoster Library

Purpose: Due to x-ray exposure and respiratory motion, a routine planning CT is usually captured with a higher in-plane resolution than through-plane resolution. Such CT with a large slice thickness may affect both contouring and dose calculation in treatment planning. We propose a self-supervised learning method to improve through-plane resolution based on the knowledge learned from self-high-resolution in-plane images.


Methods: The proposed method consists of three steps: 1) train a model on transverse images to predict high-resolution (HR) sagittal images; 2) train a second model on transverse images to predict HR coronal images; 3) generate final prediction by fusion of the predicted images in previous two steps. We generated a low-resolution (LR) image and its corresponding HR image in the x-axis to produce the training data from the same subject. Then, we extracted image patches from a transverse plane, including a pair of LR and HR patch. To establish a one-to-one mapping between LR and HR, cycle consistent generative adversarial network (CycleGAN) was used. Two CycleGAN models were trained to perform the tasks in the x-axis and y-axis respectively. At inference, LR sagittal slices and coronal slices were fed into the two networks respectively to predict HR sagittal and coronal slices. Final HR CT was obtained by image fusion of the predicted sagittal and coronal images.


Results: The mean absolute error for 20 patients was reduced from 39.8±5.0 to 23.2±4.9HU as compared to bicubic interpolation. The peak signal-noise-ratio was increased from 25.4±2.7 to 34.5±2.8dB. The normalized cross-correlation was increased from 0.962±0.006 to 0.997±0.001. All the improvements are statistically significant.


Conclusion: We have developed a novel deep learning-based self-high-resolution method to generate patient-specific HR CT, and demonstrated its feasibility and accuracy. The proposed method has great potential in improving radiation dose calculation and delivery accuracy, and decreasing CT exposure for patients.

Keywords

Not Applicable / None Entered.

Taxonomy

Not Applicable / None Entered.

Contact Email