Fig 1.
The definitions of slice thickness and scanning interval.
Illustration of the relationship between CT ‘slice thickness’ and ‘scan interval’. (Left): Set-up depicting 1 mm ‘slice thickness’ and an identical 1 mm ‘scan interval’. (Center): Configuration illustrating 3 mm ‘slice thickness’ and an identical 3 mm ‘scan interval’. (Right): Scheme showcasing a 3 mm ‘slice thickness’ with a varied 1.5 mm ‘scan interval’. Green arrows highlight the ‘slice thickness’, while red arrows indicate ‘scan intervals’. ‘Slice thickness’ defines the axial resolution, while ‘scanning interval’ refers to the gap between two consecutive slices.
Fig 2.
(A) It represents a generative model that generates a fake image from an input image (x). ‘x’ is the CT 3 mm input image. (B) It was a discriminative model that distinguishes the generated image into real and fake. ‘y’ is the 1 mm correct image. G, generative model; D, discriminative model; G(x), generated 1 mm CT image.
Fig 3.
This flow chart illustrates the process of converting N sheets of 3 mm CT images into their corresponding 1 mm counterparts using two methods within a cGAN model. Each 3 mm CT image is processed individually and the operation is repeated N times. In Method 1 (A), the i-th 3 mm CT image’s pixel value is read and input into the cGAN model in the range of [–1,1]. The model then generates an output that is scaled from the range [–1,1] to fit the CT image pixel value range of [0,4095]. Method 2 (B) also starts by reading the i-th 3 mm CT image’s pixel value and inputting it into the cGAN model in the range of [–1,1]. However, this method involves an additional step where the scaled input value [0,1] is added to the output value [–1,1] of the generation model. The final step in both methods involves determining whether the 1 mm slices at positions 2P, 1P, C, 1N, and 2N, relative to the original 3 mm slice, have been created successfully. The generated 1 mm CT image is saved, and the process concludes when there are no more 3 mm CT images to be processed.
Fig 4.
Difference image between 3 mm CT image and 1 mm CT corresponding image.
(A) 3 mm input image. (B) the corresponding 1 mm scan image, and (C) is the difference image obtained by subtracting the 1mm image from 3 mm. The red pixel value means a positive value (3 mm is a bright pixel) and the blue pixel value is representing a negative value (1 mm is the brightest pixel). The difference between the 3mm image and the corresponding 1 mm CT image is the smallest in 1 prev. images and 1 mm CT. Otherwise, the difference between the adjacent before and after images is evident in 2 prev. and 2 next. images. Prev, previous.
Fig 5.
Configuration of 3 mm CT image input and multiple 1 mm generative models (A) and the example of slice location mapping table (B). The grey column within the 3 mm image means the matched area on the 1 mm image. The 1 mm image corresponding to 3 mm was used in the Typo, and 1 mm images adjacent to 3 mm were used for training the 2P, 1P, 1N, and 2N models, respectively. C, corresponding; 2P, before 2 slices; 1P, before 1 slice; 1N, after 1 slice; 2N, after 2 slices.
Fig 6.
Two models for generating 1 mm slice CT images from 3 mm images: (A) Method 1, (B) Method 2. (A) ‘Method 1’ used a 3 mm CT raw image (x) as the input image and was configured to generate a 1 mm CT raw image (G(x)). (B) ‘Method 2’ used 3 mm CT raw image (x) as the input image, and the difference values between 3 mm and 1 mm CT images (*) were trained. The different value was added to create 1 mm CT images (+). X, CT 3 mm input image; G, generative model; G(x), generated results.
Table 1.
The results for reproducibility evaluation of training data.
Table 2.
The results for quantitative evaluation of test data.
Table 3.
The results for performance comparison of two methods on training data.
Table 4.
The results for performance comparison of two methods on test data.
Fig 7.
The evaluated sample of the set of images: (A) 3 mm CT image, (B) 1 mm CT image, (C) Image from Method 1, and (D) Image from Method 2.
Fig 8.
Quality evaluation for generated images from method 1 and 2: (A) grading for each image, and (B) the choice for better quality of image from method 1 and 2.