Quantitative Evaluation and Feasibility Study of Using AI Denoising Technique to Minimize Image Dose of 2.5MV Beam
H Kuo1*, S Lim1, S Lin2, J Sillanpaa1, L Cervino1, (1) Memorial Sloan Kettering Cancer Center, New York, NY (2) Norwalk Hospital, Norwalk, CT
Presentations
PO-GePV-M-186 (Sunday, 7/10/2022) [Eastern Time (GMT-4)]
ePoster Forums
Purpose: To evaluate image degradation of extra low dose 2.5MV beam and the feasibility of using artificial intelligent to improve the degradation.
Methods: Convolutional neural network autoencoder (CAE) was developed to remove noise from low dose (1cGy, 0.5cGy, 0.25cGy, 0.05cGy, 0.025cGy) 2.5MV image and to restore image quality to the level of higher dose (2cGy) beam. A total of 1,188 2.5MV images of head and body taken under different gantry angles (0° to 170°, 10° apart) at 6 different couch positions were used to train denoising model. Using 2cGy images as reference, peak signal to noise ratio (PSNR) and structural similarity index measure (SSIM) were calculated to evaluate the degree of image degradation of the lower dose image before and after the noise reduction from applying CAE. 12 (head and body each) 2D/3D registration of orthogonal paired KV/2.5MV beam were tested the impact of low dose (0.5cGy & 1cGy) on the setup accuracy in comparing to setup results from CBCT.
Results: Both PSNR and SSIM showed 2.5MV didn’t seriously degrade head image (<10dB) at low output till 0.25cGy, however, 15dB and 20dB (PSNR) degradation were seen at body image of 0.5cGy and 0.25cGy, respectively. CAE Model could restore image by 5-10 dB if the degradation was more than 15dB, but it added no value if degradation was <10dB. Clinical mode showed KV/2.5MV using low (1cGy & 0.5cGy), the difference of registration accuracy deviation from 3D registration using CBCT was minimal (<0.5mm/0.5°). LINAC could not be stably calibrated the 2.5MV to extra low output (0.1cGy) so that the improvement of setup accuracy using denoise image with extra low output still need to be tested.
Conclusion: CAE improves image quality of 2.5MV beam with output lower than 0.5cGy and has the potential to minimize 2.5MV image dose to low mGy.
Keywords
SNR, Noise Reduction, Registration
Taxonomy
Not Applicable / None Entered.
Contact Email