Reducing the risk of hallucinations with interpretable deep learning models for low-dose CT denoising: comparative performance analysis.

Objective.Reducing CT radiation dose is an often proposed measure to enhance patient safety, which, however results in increased image noise, translating into degradation of clinical image quality. Several deep learning methods have been proposed for low-dose CT (LDCT) denoising. The high risks posed by possible hallucinations in clinical images necessitate methods which aid the interpretation of deep learning networks. In this study, we aim to use qualitative reader studies and quantitative radiomics studies to assess the perceived quality, signal preservation and statistical feature preservation of LDCT volumes denoised by deep learning. We aim to compare interpretable deep learning methods with classical deep neural networks in clinical denoising performance.Approach.We conducted an image quality analysis study to assess the image quality of the denoised volumes based on four criteria to assess the perceived image quality. We subsequently conduct a lesion detection/segmentation study to assess the impact of denoising on signal detectability. Finally, a radiomic analysis study was performed to observe the quantitative and statistical similarity of the denoised images to standard dose CT (SDCT) images.Main results.The use of specific deep learning based algorithms generate denoised volumes which are qualitatively inferior to SDCT volumes(p< 0.05). Contrary to previous literature, denoising the volumes did not reduce the accuracy of the segmentation (p> 0.05). The denoised volumes, in most cases, generated radiomics features which were statistically similar to those generated from SDCT volumes (p> 0.05).Significance.Our results show that the denoised volumes have a lower perceived quality than SDCT volumes. Noise and denoising do not significantly affect detectability of the abdominal lesions. Denoised volumes also contain statistically identical features to SDCT volumes.

Focal unspecific bone uptake on [18F]-PSMA-1007 PET: a multicenter retrospective evaluation of the distribution, frequency, and quantitative parameters of a potential pitfall in prostate cancer imaging.

PURPOSE: Improved logistics and availability led to a rapid increase in the use of [18F]-PSMA-1007 for prostate cancer PET imaging. Initial data suggests increased uptake in benign lesions compared to [68 Ga]-PSMA-11, and clinical observations found increased unspecific bone uptake (UBU). We therefore investigate the frequency and characteristics of UBU in [18F]-PSMA-1007 PET. METHODS: We retrospectively analyzed [18F]-PSMA-1007 PET scans from four centers for the presence of UBU, defined as a focal mild-to-moderate uptake (SUVmax < 10.0) not obviously related to a benign or malignant cause. If present, up to three leading UBUs were quantified (SUVmax), localized, and correlated to clinical parameters, such as age, PSA, injected dose, Gleason score, tumor size (T1-T4), and type of PET scanner (analog vs. digital). Additionally, clinical and imaging follow-up results and therapeutic impact were evaluated. RESULTS: UBUs were identified in 179 out of 348 patients (51.4%). The most frequent localizations were ribs (57.5%) and pelvis (24.8%). The frequency of UBUs was not associated with PSA, Gleason score, tumor size, age, or the injected [18F]-PSMA-1007 dose. UBUs were significantly more frequent in images obtained with digital PET/CT scans (n = 74, 82%) than analog PET/CT scans (n = 221, 40.3%) (p = .0001) but not in digital PET/MR (n = 53, 51%) (p = .1599). In 80 out of 179 patients (44.7%), the interpretation of UBUs was critical for therapeutic management and therefore considered clinically relevant. For 65 UBUs, follow-ups were available: three biopsies, three radiotherapies with PSA follow-up, and 59 cases with imaging. After follow-up, UBUs were still considered unclear in 28 of 65 patients (43%), benign in 28 (43%), and malignant in nine (14%) patients. CONCLUSION: UBUs occur in two-thirds of patients imaged with [18F]-PSMA-1007 PET/CT and are significantly more frequent on digital PET scanners than analog scanners. UBUs should be interpreted carefully to avoid over-staging.

Impact of metal implants on xSPECT/CT Bone reconstruction: the "shining metal artefact".

BACKGROUND: Novel reconstruction algorithms, such as xSPECT Bone, are gaining more and more importance in Nuclear Medicine. With xSPECT Bone, the reconstructed emission image is enhanced by the information obtained in the corresponding CT image. The CT defines tissue classes according to the Hounsfield units. In the iterative reconstruction, each tissue class is handled separately in the forward projection step, and all together in the back projection step. As a consequence, xSPECT Bone reconstruction generates images with improved boundary delineation and better anatomic representation of tracer activity. Applying this technique, however, showed that artefacts may occur, when no uptake regions, like metal implants, exhibit fictitious uniform tracer uptake. Due to limitations in spatial resolution in gamma cameras, the xSPECT Bone reconstructed image resulted in spill-out activity from surrounding high uptake region being uniformly distributed over the metal implants. This new technology of xSPECT Bone reconstruction in general enhances the image quality of SPECT/CT; however, the potential introduction of specific artefacts which inadvertently come along with this new technology and their frequency have not yet been addressed in the literature. Therefore, the purpose of this work was to identify and characterize these specific metal artefacts (the so-called shining metal artefact) in order to reduce false positives and avoid potentially misdiagnosing loosened or infected implants. CASE PRESENTATION: In this work, we report five cases imaged with bone SPECT/CT of 5 anatomical regions (foot, elbow, spine, shoulder, ribs and knee). All cases demonstrated "shining metal artefacts" in xSPECT Bone reconstruction. CONCLUSION: While xSPECT Bone reconstruction algorithm significantly improves image quality for the diagnosis of bone and joint disorders with SPECT/CT, specific "shining metal artefacts" caused by the xSPECT Bone have to be recognized in order to avoid image misinterpretation suggesting metallic implant loosening or possible infection. The simultaneous analysis of conventionally reconstructed SPECT images (for Siemens the Flash3D reconstruction) helps to avoid misinterpretation of potential artefacts introduced by xSPECT Bone reconstruction.

Accuracy of non-contrast PMCT for determining cause of death.

The aim of this study was (1) to compare levels of accuracy regarding the categorization of causes of death between non-contrast post-mortem computed tomography (PMCT) and the final forensic report as well as between autopsy and the final forensic report, and (2) to assess levels of confidence regarding the categorization of causes of death after non-contrast PMCT and after autopsy. This prospective study was conducted over a 5 month period during which 221 cases were admitted to our institute for forensic investigations. Whole-body PMCT and forensic autopsy were performed in every case. Of these, 101 cases were included in the final study population. Inclusion criteria were: (1) age > 18 years, (2) presence of at least one of the two principal investigators at the time of admission. One radiologist and one forensic pathologist independently read all PMCT datasets using a report template. Cause of death category and confidence levels were determined by consensus. Forensic autopsy was performed by two forensic pathologists; both unblinded to imaging results. Both post-imaging and post-autopsy cause of death categorization were compared against the final cause of death, as stated in the forensic expert report, which included findings from histology and/or toxicology. Accuracy of post-imaging cause of death categorization in reference to the final cause of death category was substantial (82%, 83/101 cases, Kappa 0.752). Accuracy of post-autopsy cause of death categorization in reference to the final cause of death category was near perfect (89%, 90/101 cases, Kappa 0.852). Post-imaging sensitivity and specificity regarding the categorization of causes of death were 82% and 97%, respectively. Post-autopsy sensitivity and specificity regarding the categorization of causes of death were 89% and 98%, respectively. There was a high consistency between the accuracy of post-imaging cause of death categorization and post-imaging levels of confidence. There was less consistency between accuracy of post-autopsy cause of death categorization and post-autopsy levels of confidence. In this study categorization of causes of death based on non-contrast enhanced PMCT alone, and on PMCT and macroscopic autopsy together, proved to be consistent with the final cause of death-category as determined based on all available information including PMCT, autopsy, and (if available) histology and/or toxicology in more than 82% and 89% of all cases, respectively. There was higher consistency between levels of confidence and accuracy of causes of death categorization was higher post-imaging than post-autopsy. These results underline the fact that the diagnostic potential of PMCT goes beyond the assessment of trauma cases.

Material differentiation in forensic radiology with single-source dual-energy computed tomography.

The goal of this study was to investigate the use of dual-energy computed tomography (CT) in differentiating frequently encountered foreign material on CT images using a standard single-source CT scanner. We scanned 20 different, forensically relevant materials at two X-Ray energy levels (80 and 130 kVp) on CT. CT values were measured in each object at both energy levels. Intraclass correlation coefficient (ICC) was used to determine intra-reader reliability. Analysis of variance (ANOVA) was performed to assess significance levels between X-Ray attenuation at 80 and 130 kVp. T test was used to investigate significance levels between mean HU values of individual object pairings at single energy levels of 80 and 130 kVp, respectively. ANOVA revealed that the difference in attenuation between beam energies of 80 kVp compared to 130 kVp was statistically significant (p < 0.005) for all materials except brass and lead. ICC was excellent at 80 kVp (0.999, p < 0.001) and at 130 kVp (0.998, p < 0.001). T test showed that using single energy levels of 80 and 130 kVp respectively 181/190 objects pairs could be differentiated from one another based on HU measurements. Using the combined information from both energy levels, 189/190 object pairs could be differentiated. Scanning with different energy levels is a simple way to apply dual-energy technique on a regular single-energy CT and improves the ability to differentiate foreign bodies with CT, based on their attenuation values.

How reliable are Hounsfield-unit measurements in forensic radiology?

OBJECTIVE: To assess the reliability of computed tomography (CT) numbers, also known as Hounsfield-units (HU) in the differentiation and identification of forensically relevant materials and to provide instructions to improve the reproducibility of HU measurements in daily forensic practice. MATERIALS AND METHODS: We scanned a phantom containing non-organic materials (glass, rocks and metals) on three different CT scanners with standardized parameters. The t-test was used to assess the influence of the scanner, the size and shape of different types of regions-of-interest (ROI), the composition and shape of the object, and the reader performance on HU measurements. Intra-class correlation coefficient was used to assess intra- and inter-reader reliability. RESULTS: HU values did not change significantly as a function of ROI-shape or -size (p>0.05). Intra-reader reliability reached ICC values >0.929 (p<0.001). Inter-reader reliability was also excellent with an ICC of 0.994 (p<0.001). Four of seven objects yielded significantly different CT numbers at different levels within the object (p<0.05). In 6/7 objects the HU changed significantly from CT scanner to CT scanner (p<0.05). CONCLUSION: Reproducible CT number measurements can be achieved through correct ROI-placement and repeat measurements within the object of interest. However, HU may differ from CT-scanner to CT-scanner. In order to obtain comparable CT numbers we suggest that a dedicated Forensic Reference Phantom be developed.

Death by biscuit--exhumation, post-mortem CT, and revision of the cause of death one year after interment.

We describe a case of exhumation, performed to investigate the circumstances and cause of death, one year after burial. Post mortem computed tomography (PMCT) revealed a mass in the pharynx. Imaging directed the subsequent forensic autopsy to careful retrieval of a foreign body. Histological analysis revealed a non-cellular composition. The detection of foreign material in the pharynx and its composition indicated accidental, rather than natural death, secondary to choking on food. This unusual case illustrates how post mortem imaging can significantly contribute to forensic investigation and stresses the importance of interdisciplinary collaboration between forensic pathologists and radiologists.