This study, in which lesions were randomly simulated into the liver tissue using clinical raw data from SPECT acquisitions, demonstrated that the location of the simulated lesions is highly dependent on the local TNC ratio value and the noise level. The lesions that were detectable for the observer had statistically significant higher TNC ratios and SNR. Using MC-based OSEM reconstruction, the spatial resolution was improved [15] and consequently the TNC ratios, but the noise level remained high, influencing the SNR values.
For all reconstruction methods, the variation in liver activity concentration and high noise level resulted in disappearance of lesions when the randomly selected location was in an area of low activity concentration, about the size of the control VOI. These lesions received a lower TNC ratio due to the low activity concentration of the control VOI, and there was a higher activity concentration outside of the control VOI. The resolution volume was about the size of the control VOI, and the partial volume effect consequently spread the lesion over the area with low activity concentration. These lesions had low TNC ratios, close to one, and the visual retrospective assessment, with knowledge of the correct coordinates, resulted in the decision to consider these patients negative for lesions. This result demonstrates that not only the SPECT resolution and noise level but also the anatomical location of lesions might influence the detectability of lesions.
Since several lesions were considered negative after the retrospective analysis, we also performed an analysis of only the lesions possible to detect. Our viewpoint was that this scenario more fairly demonstrates the differences between the three reconstruction methods. The hidden lesions were naturally not detected by the observer in any of the reconstruction methods in scenario AL. Therefore, the difference between the reconstruction methods in scenarios AL and DL regarding the detection of lesions was only a reflection of the lower number of patients that are lesion positive (26 for each set in AL vs. 19 in DL). There were a larger number of misplaced lesions in scenario AL than in scenario DL (Table 7). Therefore, the ROC analysis differed more from the results of the detection of lesions and the accuracy of the assessments in scenario AL than in scenario DL. Furthermore, the misplaced lesions that had an erroneous positive impact in the ROC analysis did not influence the detection of lesions but did affect the accuracy in scenario AL. When 7 patients were changed to negative for lesion, some of these misplacements shifted (rightfully) to false positives, which explained the differences between scenarios AL and DL. Another feature of ROC analysis is that it considers the confidence level of the observer. Figure 4 shows that the confidence level was higher in filtered images compared to unfiltered images and highest in fAC OSEM. This was possibly due to the observer’s previous experience with image appearance during evaluations.
According to a previous study [15], the spatial resolution is improved in images reconstructed with MC OSEM, which would consequently imply a higher degree of detail. The perception of homogeneity of the uptake of 111In-octreotide in the liver, which is usually what the observer expects in a SPECT/CT examination, can be challenged as the spatial resolution and image quality overall improve. The blurring effect of commonly used filters also contributes to the homogenous appearance. This resembles situations in which new gamma camera designs present higher sensitivity and/or resolution. Pathologic uptake will be more intensive, but this also applies for benign findings. This might explain the higher degree of false positive findings by the observer in fMC OSEM and MC OSEM in this study, although this can also be explained by the fact that the observer claimed to be extra sensitive when identifying lesions in order to not miss them. Nevertheless, the images reconstructed with MC OSEM (specifically unfiltered) were very different visually from those most familiar to observers, which is a probable explanation for the lower confidence level reported by the observer. In this study, the observer was more confident with smoother images (Fig. 4). However, the confidence level of the observer regarding the detected lesions was highest with fMC OSEM, a probable cause of higher TNC ratios combined with low noise levels. Still, the detection of lesions and the observer accuracy were improved with unfiltered MC OSEM compared to the filtered images (Table 3).
As stated before, the ROC analysis did not consider the coordinates of the lesion, so a misplaced lesion in a patient who has a lesion was considered an accurate assessment. It can be argued that this is a correct approach when evaluating a diagnostic test for a disease, since the underlying reason for a positive test is irrelevant and the goal is to find patients who require further evaluation. In this study, however, we aimed to distinguish differences in the visual appearance of small lesions depending on the different image processing techniques. From our perspective, a misplaced lesion is due to two successive errors: the first misses the real lesion and the second finds a lesion that is not there. To consider this to be an accurate assessment is therefore unsatisfactory. Consequently, the terms detection of lesions and accuracy were chosen instead of sensitivity and specificity. As there were misplaced lesions for all reconstruction methods, and in the absence of other comparison measures, the ROC analysis was still considered valuable. This was especially true for scenario DL, where there were not as many misplaced lesions as for scenario AL.
The CT scanners used for SPECT/CT imaging at Sahlgrenska University Hospital during the years between 2004 and 2011 were, from an image quality standpoint, far inferior to the CT scanners used today. The image quality of the CT images in this study was poor, and some examinations also suffer from severe metal artifacts; these will influence the performance of the MC simulations in the reconstruction algorithm. In conventional OSEM reconstruction, the CT images are used only for attenuation correction. Hence in this study, the image quality of the CT was of less importance for fAC OSEM than for fMC OSEM and MC OSEM. Furthermore, the radial positions of the detectors at each projection angle were not registered by the gamma cameras. Therefore, the distances, in order to correct for the point spread function, had to be manually estimated (based on the CT images), which might have influenced the accuracy of the MC simulations. Hence, higher quality CT images and registered radial distances by the gamma camera, both standard in SPECT/CT scanners today, might result in more accurate MC simulations, and consequently the image quality might be further improved.
There are commercially available reconstruction methods that apply resolution recovery. At Sahlgrenska University Hospital, Gothenburg, Sweden, we have access to GE’s reconstruction method with resolution recovery, called Evolution, but in this study, we were unable to simulate lesions with Evolution. However, we previously compared the spatial resolution and showed a significant improvement in images reconstructed with MC OSEM versus Evolution [15]. Furthermore, the number of observers should ideally be more than one.
It has previously been shown that MC OSEM significantly improves the image quality in 177Lu-octreotate imaging [21] and the spatial resolution in 111In-octreotide imaging [15]. However, the noise level needs to be handled appropriately, and we aim to further investigate deep learning–generated synthetic intermediate projections (SIPs) in SPECT images, which have been demonstrated to more effectively reduce the noise level compared to post-filtering methods such as Gaussian filtering [23, 24]. This might improve SNR in images reconstructed with MC-based OSEM reconstruction.