Post-reconstruction enhancement of [¹⁸F]FDG PET images with a convolutional neural network

Ninety-seven patients referred for clinical [¹⁸F]FDG PET-CT at Skåne University Hospital, Malmö or Lund, were included in the study. Seventy-two of the patients were used for training the CNN (included December 2019 to March 2020) and a separate set of 25 patients (included April to June 2018) were used to evaluate the method. All patients underwent an intravenous injection of 4 MBq/kg body weight of [¹⁸F]FDG after at least 4 h fasting and at a glucose level ≤ 10 mM. Four Discovery MI (GE Healthcare, Milwaukee, WI, USA) PET-CT systems, each with four detector rings, were used for image acquisition. Imaging was performed 60 min after administration, and the patients were scanned from the inguinal region to the base of the skull. The patients in the training set were scanned with 5–7 bed positions, depending on the length of the patient. The acquisition time for one of the bed positions was 6 min (different bed position for different patients) and 1.5 min for the others. The images of the patients in the test group were acquired with a time per bed position of 4 min for all bed positions and stored in list-mode.

The PET-images were reconstructed using a BSREM algorithm including time-of-flight and point spread function with a 256 × 256 matrix (pixel size 2.7 × 2.7 mm², slice thickness 2.8 mm). CT images were acquired for attenuation correction and anatomic correlation of the PET images. A diagnostic CT with intravenous and oral contrast or a low-dose CT without contrast was performed.

A denoising CNN suitable for image enhancement as proposed by Zhang et al. [18] was implemented using Matlab (Mathworks Inc. Natick, Massachusetts). In our adaptation of Zhang’s model, we have changed the number of convolution layers to 10 and used a 256 × 256 × 5 (3D) matrix input. Each convolution layer consists of 68 3 × 3 × 3 filters except the first and last layer which consist of one 3 × 3 × 3 filter. As in the proposed CNN by Zhang et al., batch normalisation was used in order to speed up the training process and improve the denoising performance. Pooling layers were not used in order to benefit from a larger receptive field where contextual information can be used to improve the denoising. A linear rectifier was used after each convolution layer except the last. A mean squared error loss and a stochastic gradient descent optimizer was used. An illustration of the current paper’s adapted CNN is shown in Fig. 1.

List mode data from the bed position with a 6 min scan time were extracted from each of the 72 patients in the training group. List mode permits image reconstruction event-by-event from a chosen starting point and interval. From the list mode data, 1 image set for 6 min and 4 image sets for 1 (first, second, third and fourth minute) and 1.5 min (6 min divided in 4 intervals) were reconstructed with the BSREM algorithm, respectively. For the 1 min and 1.5 min images, a β of 500 was used, and a β of 200 was used for the 6 min image. Each reconstruction yielded 71 slices which were cropped in the beginning and the end, leaving only 50 of the centremost slices. Each 50-slice reconstruction were further divided into 10 subsets, each comprising a 3D volume sized 256 × 256 × 5, to match the input of the CNN. Furthermore, to reduce overfitting data augmentation was performed; each reconstruction (1 min and 1.5 min) was randomly resized, sheared and flipped in two dimensions, resulting in 5 additional samples. Two sets of training pairs were composed, pairs with 1 min and 6 min images and pairs with 1.5 min and 6 min images. The total training pairs for each of the two training sets were (subsets × patients × reconstructions × samples) 10 × 72 × 4 × 6 = 17.280. Two networks were trained: one for images acquired with 1 min and the other for 1.5 min/bed position. Figure 2 shows an overview how the training group and test group are set up. Table 1 shows an overview of the image sets from the training group.

25 patients (test group), each with 5 different whole-body image sets (Table 2), resulting in a total of 125 examinations, were evaluated; β500 1 min with and without CNN enhancement, β500 1.5 min with and without CNN enhancement and β300 4 min without CNN enhancement. In our department, β500 1.5 min is used clinically. The choice of β500 1 min + /− CNN is because it is a natural step if one wishes to decrease acquisition time to increase the throughput in the department. β300 4 min was considered as the gold standard image for reference. Image sets will henceforth be written without the beta-value for brevity.

The coefficient of variation (COV), considered as an objective measure of noise, was calculated from regions of interests (ROIs) drawn in the liver on transaxial images using Hermes 2.0.0 (Hermes Medical Solutions, Stockholm, Sweden). Three ROIs with a diameter of 6 cm were drawn in subsequent transaxial slices with one image in-between, and the measurements were averaged. None of the ROIs were placed where liver metastases or large vessels were seen. The ROIs were drawn in the 4 min image set and copied to the other image sets. The COV was calculated as the ratio between the SUV_{standard deviation (SD)} and the SUV_mean, thus a lower value indicates less noise.

The lesion SUV_max/peak were calculated from a VOI defined over a lesion. Three small-middle sized hotspots (pathologic lesions or physiologic uptake) per patient were selected to get a large range of SUVs. SUV_max and SUV_peak (SUV_mean in a 1 cm³ volume sphere) were calculated for all lesions.

Firstly, a nuclear medicine specialist evaluated all image sets per patient side-by-side in all anatomic planes to ensure the CNN didn’t add or subtract anything relevant from the images.

Secondly, two nuclear medicine specialists, including the previously mentioned and a radiology resident was presented with a blinded list consisting the 125 examinations in random order. Noise and contrast levels were recorded for each examination on a 5-point scale.

A training/calibration session was held to establish a baseline for noise and contrast levels. Images from patients not included in the test group with previously 5 mentioned image sets series were displayed side-by-side. 3 points in noise and contrast was established as baseline for standard 1.5 min (the clinically used images). For noise, 5 = “very little noise”, 4 = “slightly less noise compared with baseline”, 2 = “slightly increased noise compared with baseline”, 1 = markedly increased noise”. For contrast, 5 = “very good contrast”, 4 = “slightly increased contrast compared with baseline”, 2 = “slightly decreased contrast compared with baseline” and 1 = “markedly decreased contrast”.

Assessment of examinations was performed on dedicated workstations using Hermes 2.0.0.

Friedman’s test was used to compare quantitative data for SUV_max, SUV_peak and COV measurements in all 5 image sets (4 min, 1.5 min + /− CNN, 1 min + /− CNN), significant p value was set at p < 0.05. Post-hoc analysis with Wilcoxon signed-rank test was conducted with a Bonferroni correction applied, resulting in a significance level set at p < 0.005. Since there were 5 image sets, each measurement group resulted in 10 comparisons and thus significant p value was calculated as 0.05/10 = 0.005.

All statistical computations were performed using IBM SPSS Statistics version 26.0.0.1 (IBM, Armonk, NY, USA).

The differences in the groups SUV_peak, SUV_max, and COV were statistically significant. A majority of the post-hoc comparisons showed significant differences for all quantitative metrics (SUV_peak, SUV_max and COV). For SUV_peak, the differences in the pairs 1.5 min–1.5 min CNN, 1.5 min–1 min and 1.5 min CNN–1 min, respectively, fell short of significance. For SUV_max, the pairs 1.5 min–1.5 min CNN and 1.5 min CNN–1 min, respectively, fell short of significance. A graphical representation of the results is shown in Fig. 5 and statistical results in Table 3.

Reader 2 evaluated all image sets in the test group and found no added or subtracted hotspots in any image set.

Mean and SD of noise and contrast score respectively for each image set of each reader are presented in Table 4. Noise-wise, 1.5 min CNN had the highest mean score for all readers; contrast-wise, reader 1 scored 1.5 min CNN the highest followed by 4 min, whereas reader 2 and 3 scored 4 min the highest followed by 1.5 min CNN. The standard 1 min acquisition got lowest scores from all readers both regarding noise and contrast. For all readers, 1.5 min CNN and 1 min CNN were consistently ranked higher than their corresponding 1.5 min and 1 min image set. Rankings of image sets for noise and contrast are shown in Table 5a, b.

Group wise analysis showed significant differences in scoring of the image sets for noise or contrast for any reader (p < 0.001). In the post-hoc pairwise analysis we found that there were significant differences in scoring for both noise and contrast for 1.5–1.5 min CNN across all comparisons. In the 1.5–1 min CNN comparisons, half of the mean rank scores were significantly different, detailed results for all pairwise comparisons are shown in Table 6.