핵의학 갑상샘 팬텀 영상에서 보간법이 초고분해능 ResNet 모델성능에 미치는 영향에 관한 연구 (Study on Impact of Interpolation Methods on Performance of Super-Resolution ResNet Model in Nuclear Medicine Thyroid Phantom Imaging)

서지정보

· 발행기관 : 대한방사선과학회
· 수록지 정보 : Journal of Radiological Science and Technology / 47권 / 6호 / 433 ~ 439페이지
· 저자명 : 김민하, 박찬록

초록

The nuclear medicine imaging can improve the image quality through the application of various interpolation techniques. Additionally, deep learning algorithm, which perform feature extraction between input and label datasets, is widely utilized to improve image quality in nuclear medicine. Thus, the purpose of this study was to confirm the performance of deep learning algorithm according to applied various interpolation methods as input data using the thyroid phantom images. A total of 200 thyroid phantom images, each sized 256 × 256, were acquired at an activity of 37 M㏃ for 1 minute to generate the label images. Interpolation methods including nearest neighbor, bilinear, biquadratic, bicubic, biquartic, and biquintic were applied during both downsampling and upsampling processes. The super-resolution residual network (SRResNet) architecture was implemented with a learning rate of 0.0001 and 300 epochs, using an 8:1:1 ratio for train, validation, and test sets, respectively. The generated output images analyzed using coefficient of variation (COV) and contrast to noise ratio (CNR). Consequently, the SRResNet algorithm, which used low-resolution images generated with the bicubic interpolation method, showed the highest performance. This study demonstrates the importance of selecting appropriate interpolation methods for generating input images to improve the accuracy of the SRResNet algorithm in nuclear medicine thyroid imaging and even in other medical fields for diagnosis.

영어초록

The nuclear medicine imaging can improve the image quality through the application of various interpolation techniques. Additionally, deep learning algorithm, which perform feature extraction between input and label datasets, is widely utilized to improve image quality in nuclear medicine. Thus, the purpose of this study was to confirm the performance of deep learning algorithm according to applied various interpolation methods as input data using the thyroid phantom images. A total of 200 thyroid phantom images, each sized 256 × 256, were acquired at an activity of 37 M㏃ for 1 minute to generate the label images. Interpolation methods including nearest neighbor, bilinear, biquadratic, bicubic, biquartic, and biquintic were applied during both downsampling and upsampling processes. The super-resolution residual network (SRResNet) architecture was implemented with a learning rate of 0.0001 and 300 epochs, using an 8:1:1 ratio for train, validation, and test sets, respectively. The generated output images analyzed using coefficient of variation (COV) and contrast to noise ratio (CNR). Consequently, the SRResNet algorithm, which used low-resolution images generated with the bicubic interpolation method, showed the highest performance. This study demonstrates the importance of selecting appropriate interpolation methods for generating input images to improve the accuracy of the SRResNet algorithm in nuclear medicine thyroid imaging and even in other medical fields for diagnosis.

참고자료

· 없음