Purpose: We investigate how an intrinsic speckle tracking approach to speckle-based x-ray imaging is used to extract an object’s effective dark-field (DF) signal, which is capable of providing object information in three dimensions.

Approach: The effective DF signal was extracted using a Fokker–Planck type formalism, which models the deformations of illuminating reference beam speckles due to both coherent and diffusive scatter from the sample. Here, we assumed that (a) small-angle scattering fans at the exit surface of the sample are rotationally symmetric and (b) the object has both attenuating and refractive properties. The associated inverse problem of extracting the effective DF signal was numerically stabilized using a “weighted determinants” approach.

Results: Effective DF projection images, as well as the DF tomographic reconstructions of the wood sample, are presented. DF tomography was performed using a filtered back projection reconstruction algorithm. The DF tomographic reconstructions of the wood sample provided complementary, and otherwise inaccessible, information to augment the phase contrast reconstructions, which were also computed.

Conclusions: An intrinsic speckle tracking approach to speckle-based imaging can tomographically reconstruct an object’s DF signal at a low sample exposure and with a simple experimental setup. The obtained DF reconstructions have an image quality comparable to alternative x-ray DF techniques.

Purpose: Surgery is an essential part of the curative plan for most patients affected with solid tumors. The outcome of such surgery, e.g., recurrence rates and ultimately patient survival, depends on several factors where the resection margin is of key importance. Presently, the resection margin is assessed by classical histology, which is time-consuming (several days), destructive, and basically only gives two-dimensional information. Clearly, it would be advantageous if immediate feedback on tumor extension in all three dimensions were available to the surgeon intraoperatively.

Approach: We investigate a laboratory propagation-based phase-contrast x-ray computed tomography system that provides the resolution, the contrast, and, potentially, the speed for this purpose. The system relies on a liquid-metal jet microfocus source and a scintillator-coated CMOS detector. Our study is performed on paraffin-embedded non-stained samples of human pancreatic neuroendocrine tumors, liver intrahepatic cholangiocarcinoma, and pancreatic serous cystic neoplasm (benign).

Results: We observe tumors with distinct and sharp edges having cellular resolution (∼10  μm) as well as many assisting histological landmarks, allowing for resection margin assessment. All x-ray data are compared with classical histology. The agreement is excellent.

Conclusion: We conclude that the method has potential for intraoperative three-dimensional virtual histology.

Purpose: Tomography using diffracted x-rays produces reconstructions mapping quantities such as crystal lattice parameter(s), crystallite size, and crystallographic texture, information quite different from that obtained with absorption or phase contrast. Diffraction tomography is used to map an entire blue shark centrum with its double cone structure (corpora calcerea) and intermedialia (four wedges).

Approach: Energy dispersive diffraction (EDD) and polychromatic synchrotron x-radiation at 6-BM-B, the Advanced Photon Source, were used. Different, properly oriented Bragg planes diffract different x-ray energies; these intensities are measured by one of ten energy-sensitive detectors. A pencil beam defines the irradiated volume, and a collimator before each energy-sensitive detector selects which portion of the irradiated column is sampled at any one time. Translating the specimen along X  ,  Y, and Z axes produces a 3D map.

Results: We report 3D maps of the integrated intensity of several bioapatite reflections from the mineralized cartilage centrum of a blue shark. The c axis reflection’s integrated intensities and those of a reflection with no c axis component reveal that the cone wall’s bioapatite is oriented with its c axes lateral, i.e., perpendicular to the backbone’s axis, and that the wedges’ bioapatite is oriented with its c axes axial. Absorption microcomputed tomography (laboratory and synchrotron) and x-ray excited x-ray fluorescence maps provide higher resolution views.

Conclusion: The bioapatite in the cone walls and wedges is oriented to resist lateral and axial deflections, respectively. Mineralized tissue samples can be mapped in 3D with EDD tomography and subsequently studied by destructive methods.

Purpose: Dental calculus forms on teeth during the life of an individual and its investigation can yield information about diet, health status, and environmental pollution. Currently, the analytical techniques used to visualize the internal structure of human dental calculus and entrapped inclusions are limited and require destructive sampling, which cannot always be justified.

Approach: We used propagation phase-contrast synchrotron radiation micro-computed tomography (PPC-SR-μCT) to non-destructively examine the internal organization of dental calculus, including its microstructure and entrapped inclusions, on both modern and archeological samples.

Results: The virtual histological exploration of the samples shows that PPC-SR-μCT is a powerful approach to visualize the internal organization of dental calculus. We identified several important features, including previously undetected negative imprints of enamel and dentine growth markers (perikymata and periradicular bands, respectively), the non-contiguous structure of calculus layers with multiple voids, and entrapped plant remains.

Conclusions: PPC-SR-μCT is an effective technique to explore dental calculus structural organization, and is especially powerful for enabling the identification of inclusions. The non-destructive nature of synchrotron tomography helps protect samples for future research. However, the irregular layers and frequent voids reveal a high heterogeneity and variability within calculus, with implications for research focusing on inclusions.

Purpose: Propagation-based x-ray imaging (PBI) is a phase-contrast technique that is employed in high-resolution imaging by introducing some distance between sample and detector. PBI causes characteristic intensity fringes that have to be processed with appropriate phase-retrieval algorithms, which has historically been a difficult task for objects composed of several different materials. Spectral x-ray imaging has been introduced to PBI to overcome this issue and to potentially utilize the spectral nature of the data for material-specific imaging. We aim to explore the potential of spectral PBI in three-dimensional computed tomography (CT) imaging in this work.

Approach: We demonstrate phase-retrieval for experimental high-resolution spectral propagation-based CT data of a simple two-component sample, as well as a multimaterial capacitor test sample. Phase-retrieval was performed using an algorithm based on the Alvarez–Macovski model. Virtual monochromatic (VMI) and effective atomic number images were calculated after phase-retrieval.

Results: Phase-retrieval results from the spectral data set show a distinct gray-level for each material with no residual phase-contrast fringes. Several representations of the phase-retrieved data are provided. The VMI is used to display an attenuation-equivalent image at a chosen display energy of 80 keV, to provide good separation of materials with minimal noise. The effective atomic number image shows the material composition of the sample.

Conclusions: Spectral photon-counting detector technology has already been shown to be compatible with spectral PBI, and there is a foreseeable need for robust phase-retrieval in high-resolution, spectral x-ray CT in the future. Our results demonstrate the feasibility of phase-retrieval for spectral PBI CT.

Purpose: Synchrotron radiation-based tomography yields microanatomical features in human and animal tissues without physical slicing. Recent advances in instrumentation have made laboratory-based phase tomography feasible. We compared the performance of three cutting-edge laboratory systems benchmarked by synchrotron radiation-based tomography for three specimens. As an additional criterion, the user-friendliness of the three microtomography systems was considered.

Approach: The three tomography systems—SkyScan 2214 (Bruker-microCT, Kontich, Belgium), Exciscope prototype (Stockholm, Sweden), and Xradia 620 Versa (Zeiss, Oberkochen, Germany)—were given 36 h to measure three medically relevant specimens, namely, zebrafish larva, archaeological human tooth, and porcine nerve. The obtained datasets were registered to the benchmark synchrotron radiation-based tomography from the same specimens and selected ones to the SkyScan 1275 and phoenix nanotom m^® laboratory systems to characterize development over the last decade.

Results: Next-generation laboratory-based microtomography almost reached the quality achieved by synchrotron-radiation facilities with respect to spatial and density resolution, as indicated by the visualization of the medically relevant microanatomical features. The SkyScan 2214 system and the Exciscope prototype demonstrated the complementarity of phase information by imaging the eyes of the zebrafish larva. The 3-μm thin annual layers in the tooth cementum were identified using Xradia 620 Versa.

Conclusions: SkyScan 2214 was the simplest system and was well-suited to visualizing the wealth of anatomical features in the zebrafish larva. Data from the Exciscope prototype with the high photon flux from the liquid metal source showed the spiral nature of the myelin sheaths in the porcine nerve. Xradia 620 Versa, with detector optics as typically installed for synchrotron tomography beamlines, enabled the three-dimensional visualization of the zebrafish larva with comparable quality to the synchrotron data and the annual layers in the tooth cementum.

Purpose: Existing maximum-likelihood (ML) methods in computed tomography usually require significant computing resources to implement, and/or are limited to particular measurement noise models that are representative of the simplest theoretical archetypes. There is an absence of general procedures to produce rapid ML methods that account precisely for the noise model of a given experiment. We investigate a mathematical-computational procedure of producing constrained quadratic optimization reconstruction algorithms that fill this niche, requiring less computing resources than the exact (expectation-maximization) procedures and having comparable performance with least-squares iterative methods. This allows high-fidelity reconstructions to be practically achievable for largely arbitrary noise models.

Approach: We identify a systematic mathematical procedure to produce constrained quadratic optimization methods that maximize tomogram likelihood under arbitrary noise models, which are tunable to specific characteristics of the experiment. This procedure is applied to a general theory of mixed Poisson–Gaussian noise in transmission tomography, and to a theory of invertible linear transformations of measurement intensity subject to Poisson noise. We perform tomographic reconstructions of a very highly attenuating two-dimensional object phantom and compare the speed and fidelity of reconstruction with alternative quadratic metrics (ℓ²—minimization among others).

Results: Quantitative metrics reveal that reconstructions under our systematically produced quadratic methods achieved significantly greater reconstruction fidelity with less computation than the optimized conventional, untuned quadratic metrics with a comparable procedure.

Conclusion: Constrained quadratic optimization methods appear to apply sufficiently good approximations to achieve a high reconstruction fidelity with a simple quadratic metric amenable to a broad class of minimization methods. These preliminary simulation-based results are very promising and suggest that such methods may be used to produce high-fidelity reconstructions with less computation than many other statistical methods. By design, these quadratic methods are also explicit and quantitative in their description, allowing fine-tuning according to the specific uncertainties and noise model of the experiment. Further research is required to ascertain the full practical potential of these methods.

Purpose: The morphology of a polymer aligner, designed according to an orthodontic treatment plan, determines clinical outcomes. A fundamental element of orthodontic tooth movement with aligner treatment is the fit of the aligner’s surface to the individual teeth. Gaps between the aligner and teeth do occur because current aligner fabrication is not capable of completely reproducing the complex anatomy of the individual denture. Our study aims at a quantitative three-dimensional assessment of the fit between optically transparent aligners placed on a polymeric model of the upper dental arch for two thermofoil thicknesses at preselected thermoforming temperatures.

Approach: Using an intraoral scan of a subject’s upper dental arch, eight models were printed using a stereolithographic system. A series of eight NaturAligners^® was manufactured with a pressure molding process, using thermofoils with thicknesses of 550 and 750  μm and preselected process temperatures between 110°C and 210°C. These aligners placed on the corresponding models were imaged by an advanced micro computed tomography system. The aligners and the models were segmented to extract the gaps and aligners’ local thicknesses as a function of the processing temperature for the two foil thicknesses.

Results: The results indicate that the aligners show a better fit when the foils are processed at higher temperatures. Nevertheless, processing temperatures can be kept below 150°C, as the gain becomes negligible. Thermal processing reduces the average thickness of the aligners to 60% with respect to the planar starting foil. These thickness distributions demonstrate that the aligners are generally thicker on the occlusal surfaces of molars and premolars but thinner around the incisors and buccal as well as on oral surfaces.

Conclusions: Hard x-ray tomography with micrometer resolution is a powerful technique employed to localize the gaps between aligners and teeth, and it also enables film thickness measurements after thermoforming. The thicker film on the occlusal surfaces is most welcome because of aligner abrasion during wear. The NaturAligner^® surfaces consist of a 25-μm-thin cellulose layer, and thus the microplastics released via abrasion of less than this thickness are expected to be substantially less critical than for other commercially available, optically transparent aligners.

Purpose: Quantitative measures derived from positron emission tomography (PET) images are subject to statistical uncertainty, depending critically on system parameters, including the spatial resolution of the scanner. Predictions of statistical uncertainty of quantitative measures were compared with measurements.

Approach: Measurements were performed on the dual-ring PET prototype setup at the University of Michigan. The setup consisted of multiple detectors that, in combination, span a system resolution ranging between 1 and 5.5 mm full-width-at-half-maximum (FWHM). A Micro Jaszczak hot-spot phantom with rod diameters between 1.2 and 4.8 mm was imaged and independently reconstructed for different detector combinations. Statistical properties of quantitative measures were evaluated for different reconstructions.

Results: Measured signal-to-noise ratios (SNR) of 108  ±  14, 85  ±  11, and 40  ±  5 for high (0.92 to 0.98 mm FWHM), medium (1.3 to 1.5 mm FWHM), and low (5.5 mm FWHM) resolution detector configurations and 1 million events in general terms followed predications based on the detector resolution.

Conclusions: The unique tomograph configuration allowed for experimental comparison of the impact of spatial resolution on statistical properties of reconstructions in the same setup. An SNR advantage in systems with high resolutions was predicted and confirmed even for object features significantly larger than the detector resolution.

Purpose: Malignant breast lesions can be distinguished from benign lesions by their mechanical properties. This has been utilized for mechanical imaging in which the stress distribution over the breast is measured. Mechanical imaging has shown the ability to identify benign or normal cases and to reduce the number of false positives from mammography screening. Our aim was to develop a model of mechanical imaging acquisition for simulation purposes. To that end, we simulated mammographic compression of a computer model of breast anatomy and lesions.

Approach: The breast compression was modeled using the finite element method. Two finite element breast models of different sizes were used and solved using linear elastic material properties in open-source virtual clinical trial (VCT) software. A spherical lesion (15 mm in diameter) was inserted into the breasts, and both the location and stiffness of the lesion were varied extensively. The average stress over the breast and the average stress at the lesion location, as well as the relative mean pressure over lesion area (RMPA), were calculated.

Results: The average stress varied 6.2–6.5 kPa over the breast surface and 7.8–11.4 kPa over the lesion, for different lesion locations and stiffnesses. These stresses correspond to an RMPA of 0.80 to 1.46. The average stress was 20% to 50% higher at the lesion location compared with the average stress over the entire breast surface.

Conclusions: The average stress over the breast and the lesion location corresponded well to clinical measurements. The proposed model can be used in VCTs for evaluation and optimization of mechanical imaging screening strategies.

Purpose: Image-based analysis of breast tumor growth rate may optimize breast cancer screening and diagnosis by suggesting optimal screening intervals and guide the clinical discussion regarding personalized screening based on tumor aggressiveness. Simulation-based virtual clinical trials (VCTs) can be used to evaluate and optimize medical imaging systems and design clinical trials. This study aimed to simulate tumor growth over multiple screening rounds.

Approach: This study evaluates a preliminary method for simulating tumor growth. Clinical data on tumor volume doubling time (TVDT) was used to fit a probability distribution (“clinical fit”) of TVDTs. Simulated tumors with TVDTs sampled from the clinical fit were inserted into 30 virtual breasts (“simulated cohort”) and used to simulate mammograms. Based on the TVDT, two successive screening rounds were simulated for each virtual breast. TVDTs from clinical and simulated mammograms were compared. Tumor sizes in the simulated mammograms were measured by a radiologist in three repeated sessions to estimate TVDT.

Results: The mean TVDT was 297 days (standard deviation, SD, 169 days) in the clinical fit and 322 days (SD, 217 days) in the simulated cohort. The mean estimated TVDT was 340 days (SD, 287 days). No significant difference was found between the estimated TVDTs from simulated mammograms and clinical TVDT values (p  >  0.5). No significant difference (p  >  0.05) was observed in the reproducibility of the tumor size measurements between the two screening rounds.

Conclusions: The proposed method for tumor growth simulation has demonstrated close agreement with clinical results, supporting potential use in VCTs of temporal breast imaging.

Purpose: We set out a fully developed algorithm for adapting mammography images to appear as if acquired using different technique factors by changing the signal and noise within the images. The algorithm accounts for difference between the absorption by the object being imaged and the imaging system.

Approach: Images were acquired using a Hologic Selenia Dimensions x-ray unit for the validation, of three thicknesses of polymethyl methacrylate (PMMA) blocks with or without different thicknesses of PMMA contrast objects acquired for a range of technique factors. One set of images was then adapted to appear the same as a target image acquired with a higher or lower tube voltage and/or a different anode/filter combination. The average linearized pixel value, normalized noise power spectra (NNPS), and standard deviation of the flat field images and the contrast-to-noise ratio (CNR) of the contrast object images were calculated for the simulated and target images. A simulation study tested the algorithm on images created using a voxel breast phantom at different technique factors and the images compared using local signal level, variance, and power spectra.

Results: The average pixel value, NNPS, and standard deviation for the simulated and target images were found to be within 9%. The CNRs of the simulated and target images were found to be within 5% of each other. The differences between the target and simulated images of the voxel phantom were similar to those of the natural variability.

Conclusions: We demonstrated that images can be successfully adapted to appear as if acquired using different technique factors. Using this conversion algorithm, it may be possible to examine the effect of tube voltage and anode/filter combination on cancer detection using clinical images.

Purpose: To assess the result in orthopedic trauma surgery, usually three-dimensional volume data of the treated region is acquired. With mobile C-arm systems, these acquisitions can be performed intraoperatively, reducing the number of required revision surgeries. However, the acquired volumes are typically not aligned to the anatomical regions. Thus, the multiplanar reconstructed (MPR) planes need to be adjusted manually during the review of the volume. To speed up and ease the workflow, an automatic parameterization of these planes is needed.

Approach: We present a detailed study of multitask learning (MTL) regression networks to estimate the parameters of the MPR planes. First, various mathematical descriptions for rotation, including Euler angle, quaternion, and matrix representation, are revised. Then, two different MTL network architectures based on the PoseNet are compared with a single task learning network.

Results: Using a matrix description rather than the Euler angle description, the accuracy of the regressed normals improves from 7.7 deg to 7.3 deg in the mean value for single anatomies. The multihead approach improves the regression of the plane position from 7.4 to 6.1 mm, whereas the orientation does not benefit from this approach. Thus, the achieved accuracy meets the reported interrater variance in similarly complex body regions of up to 6.3 deg for the normals and up to 9.3 mm for the plane position.

Conclusions: The use of a multihead approach with shared features leads to more accurate plane regression compared with the use of individual networks for each task. It also improves the angle estimation for the ankle region. The reported results are in the same range as manual plane adjustments. The use of a combined network with shared parameters requires less memory, which is a great benefit for the implementation of an application for the surgical environment.

Purpose: To derive a multinomial probability function and quantitative measures of the data and epistemic uncertainty as direct output of a 3D U-Net segmentation network.

Approach: A set of T1 brain MRI images were downloaded from the Connectome Project and segmented using FMRIB’s FAST algorithm to be used as ground truth. A 3D U-Net neural network was trained with sample sizes of 200, 500, and 898 T1 brain images using a loss function defined as the negative logarithm of the likelihood based on a derivation of the definition of the multinomial probability function. From this definition, the epistemic and aleatoric uncertainty equations were derived and used to quantify maps of the uncertainty along with tissue segmentations.

Results: Maps of the tissue segmentation along with the epistemic and aleatoric uncertainty, per voxel, are presented. The uncertainty decreased based on the increasing number of training data used to train the neural network. The neural network trained with 898 volumes resulted in uncertainty maps that were high primarily in the tissue boundary regions. The epistemic and aleatoric uncertainty were averaged over all test data (connectome and tumor separately), and the epistemic uncertainty showed a decreasing trend, as expected, with increasing numbers of data used to train the model. The aleatoric uncertainty showed a similar trend which was also expected as the aleatoric uncertainty is not expected to be as dependent on the number of training data.

Conclusion: The derived uncertainty equations from a multinomial probability distribution were able to quantify the aleatoric and epistemic uncertainty per voxel and are applicable for all two-dimensional and three-dimensional neural networks.

Purpose: Rapid prognostication of COVID-19 patients is important for efficient resource allocation. We evaluated the relative prognostic value of baseline clinical variables (CVs), quantitative human-read chest CT (qCT), and AI-read chest radiograph (qCXR) airspace disease (AD) in predicting severe COVID-19.

Approach: We retrospectively selected 131 COVID-19 patients (SARS-CoV-2 positive, March to October, 2020) at a tertiary hospital in the United States, who underwent chest CT and CXR within 48 hr of initial presentation. CVs included patient demographics and laboratory values; imaging variables included qCT volumetric percentage AD (POv) and qCXR area-based percentage AD (POa), assessed by a deep convolutional neural network. Our prognostic outcome was need for ICU admission. We compared the performance of three logistic regression models: using CVs known to be associated with prognosis (model I), using a dimension-reduced set of best predictor variables (model II), and using only age and AD (model III).

Results: 60/131 patients required ICU admission, whereas 71/131 did not. Model I performed the poorest (AUC  =  0.67 [0.58 to 0.76]; accuracy  =  77  %  ). Model II performed the best (AUC  =  0.78 [0.71 to 0.86]; accuracy  =  81  %  ). Model III was equivalent (AUC  =  0.75 [0.67 to 0.84]; accuracy  =  80  %  ). Both models II and III outperformed model I (AUC difference  =  0.11 [0.02 to 0.19], p  =  0.01; AUC difference  =  0.08 [0.01 to 0.15], p  =  0.04, respectively). Model II and III results did not change significantly when POv was replaced by POa.

Conclusions: Severe COVID-19 can be predicted using only age and quantitative AD imaging metrics at initial diagnosis, which outperform the set of CVs. Moreover, AI-read qCXR can replace qCT metrics without loss of prognostic performance, promising more resource-efficient prognostication.

Background: Ultrasound (US)-guided fine needle aspiration (FNA) cytology is the gold standard for the evaluation of thyroid nodules. However, up to 30% of FNA results are indeterminate, requiring further testing. In this study, we present a machine-learning analysis of indeterminate thyroid nodules on ultrasound with the aim to improve cancer diagnosis.

Methods: Ultrasound images were collected from two institutions and labeled according to their FNA (F) and surgical pathology (S) diagnoses [malignant (M), benign (B), and indeterminate (I)]. Subgroup breakdown (FS) included: 90 BB, 83 IB, 70 MM, and 59 IM thyroid nodules. Margins of thyroid nodules were manually annotated, and computerized radiomic texture analysis was conducted within tumor contours. Initial investigation was conducted using five-fold cross-validation paradigm with a two-class Bayesian artificial neural networks classifier, including stepwise feature selection. Testing was conducted on an independent set and compared with a commercial molecular testing platform. Performance was evaluated using receiver operating characteristic analysis in the task of distinguishing between malignant and benign nodules.

Results: About 1052 ultrasound images from 302 thyroid nodules were used for radiomic feature extraction and analysis. On the training/validation set comprising 263 nodules, five-fold cross-validation yielded area under curves (AUCs) of 0.75 [Standard Error (SE) = 0.04; P  <  0.001] and 0.67 (SE = 0.05; P  =  0.0012) for the classification tasks of MM versus BB, and IM versus IB, respectively. On an independent test set of 19 IM/IB cases, the algorithm for distinguishing indeterminate nodules yielded an AUC value of 0.88 (SE = 0.09; P  <  0.001), which was higher than the AUC of a commercially available molecular testing platform (AUC = 0.81, SE = 0.11; P  <  0.005).

Conclusion: Machine learning of computer-extracted texture features on gray-scale ultrasound images showed promising results classifying indeterminate thyroid nodules according to their surgical pathology.

Purpose: We demonstrate continuous learning and assess its impact on the performance of artificial intelligence of breast dynamic contrast-enhanced magnetic resonance imaging in the task of distinguishing malignant from benign lesions on an independent clinical test dataset.

Approach: The study included 1979 patients with 1990 lesions who underwent breast MR imaging during 2015, 2016, and 2017, retrospectively collected under an IRB-approved protocol; there were 1494 malignant and 496 benign lesions based on histopathology. AI was conducted in the task of distinguishing malignant and benign lesions, and independent testing was performed to assess the effect of increasing the numbers of training cases. Five training sets mimicking clinical implementation of continuous AI learning included cases from (1) first quarter of 2015, (2) first half of 2015, (3) all 2015, (4) all 2015 and first half of 2016, and (5) all 2015 and 2016. All classifiers were evaluated on the 2017 independent test set. The area under the ROC curve (AUC) served as the performance metric and was calculated over all lesions in the test set, as well as only mass lesions and only non-mass enhancements. The Mann–Kendall test was used to determine if continuous learning resulted in a positive trend in classification performance. P  <  0.05 was considered to be statistically significant.

Results: Over the continuous training period, the selected feature subsets tended to become more similar and stable. Performance of the five training conditions on the independent test dataset yielded AUCs of 0.86 (95% CI: [0.83,0.90]), 0.87 (95% CI: [0.83,0.90]), 0.88 (95% CI: [0.84,0.91]), 0.89 (95% CI: [0.85,0.92]), and 0.89 (95% CI: [0.86,0.92]). The Mann–Kendall test indicated a statistically significant positive trend (P  =  0.0167) in classification performance with continuous learning.

Conclusions: Improved diagnostic performance over time was observed when continuous learning of AI was implemented on an independent clinical test dataset.

Purpose: The purpose of our study was to analyze dental panoramic radiographs and contribute to dentists’ diagnosis by automatically extracting the information necessary for reading them. As the initial step, we detected teeth and classified their tooth types in this study.

Approach: We propose single-shot multibox detector (SSD) networks with a side branch for 1-class detection without distinguishing the tooth type and for 16-class detection (i.e., the central incisor, lateral incisor, canine, first premolar, second premolar, first molar, second molar, and third molar, distinguished by the upper and lower jaws). In addition, post-processing was conducted to integrate the results of the two networks and categorize them into 32 classes, differentiating between the left and right teeth. The proposed method was applied to 950 dental panoramic radiographs obtained at multiple facilities, including a university hospital and dental clinics.

Results: The recognition performance of the SSD with a side branch was better than that of the original SSD. In addition, the detection rate was improved by the integration process. As a result, the detection rate was 99.03%, the number of false detections was 0.29 per image, and the classification rate was 96.79% for 32 tooth types.

Conclusions: We propose a method for tooth recognition using object detection and post-processing. The results show the effectiveness of network branching on the recognition performance and the usefulness of post-processing for neural network output.

Purpose: Photon counting imaging detectors (PCD) has paved the way for spectral x-ray computed tomography (spectral CT), which simultaneously measures a sample’s linear attenuation coefficient (LAC) at multiple energies. However, cadmium telluride (CdTe)-based PCDs working under high flux suffer from detector effects, such as charge sharing and photon pileup. These effects result in the severe spectral distortions of the measured spectra and significant deviation of the extracted LACs from the reference attenuation curve. We analyze the influence of the spectral distortion correction on material classification performance.

Approach: We employ a spectral correction algorithm to reduce the primary spectral distortions. We use a method for material classification that measures system-independent material properties, such as electron density, ρ_e, and effective atomic number, Z_eff. These parameters are extracted from the LACs using attenuation decomposition and are independent of the scanner specification. The classification performance with the raw and corrected data is tested on different numbers of energy bins and projections and different radiation dose levels. We use experimental data with a broad range of materials in the range of 6  ≤  Z_eff  ≤  15, acquired with a custom laboratory instrument for spectral CT.

Results: We show that using the spectral correction leads to an accuracy increase of 1.6 and 3.8 times in estimating ρ_e and Z_eff, respectively, when the image reconstruction is performed from only 12 projections and the 15 energy bins approach is used.

Conclusions: The correction algorithm accurately reconstructs the measured attenuation curve and thus gives better classification performance.

Purpose: This work aims to develop a framework to accurately and efficiently simulate metallic objects used during interventional oncology (IO) procedures and their artifacts in computed tomography (CT) images of different body regions.

Approach: A metal insertion framework based on an existing lesion insertion tool was developed. Noise and beam hardening models were incorporated into the model and validated by comparing images of real and artificially inserted metallic rods of known material composition and dimensions. The framework was further validated by inserting ablation probes into a water phantom and comparing image appearance to scans of real probes at matching locations in the phantom. Finally, a comprehensive library of metallic probes used in our IO practice was generated and a graphical user interface was built to efficiently insert any number of probes at arbitrary positions in patient CT data, including projection and image domain insertions.

Results: Metallic rod experiments demonstrated that noise and beam hardening were properly modeled. Phantom and patient data with virtually inserted probes demonstrated similar artifact appearance and magnitude compared with real probes. The developed user interface resulted in accurately co-registered virtual probes both with and without accompanying artifacts from projection and image domain insertions, respectively.

Conclusions: The developed metal insertion framework successfully replicates metallic object and artifact appearance with projection domain insertions and provides corresponding artifact-free images with the metallic object in the identical location through image domain insertion. This framework has potential to generate robust training libraries for deep learning algorithms and facilitate image quality optimization in interventional CT.

Purpose: The introduction of whole slide imaging and digital pathology has enabled greater scrutiny of visual search behaviors among pathologists. We aim to investigate zooming and panning behaviors, external markers of visual processing capabilities, and the changes with experience.

Approaches: Twenty digitized breast core needle biopsy histopathology slides were obtained from the circulating slides from the main digital pathology trial (IRAS number: 258799). These were presented to five pathologists with varying experience (1.5 to 40 years) whose examinations were recorded. Data of visual fixations were collected using eye-tracking cameras, and the magnification data and zooming behaviors were extracted in an objective fashion by an automated algorithm. The relationship between experience and metrics was analyzed using mixed-effects regression analyses.

Results: There was a significant association between experience and both reading times (p  <  0.001) and a number of fixations (p  <  0.001), with these relationships being inversely proportional. The greater experience was also associated with greater diagnostic accuracy (p  =  0.033). We found that experience was significantly associated with greater use of magnification changes (p  <  0.001). Conversely, less experience showed a near significant association with the increased proportion of time spent panning (p  =  0.070).

Conclusions: Fewer fixations needed to reach a diagnosis and quicker reading times are indicative of greater cognitive and visual processing capabilities with greater experience. These cognitive capabilities may be a prerequisite for the more frequent zooming changes that are more prevalent with increasing experience.

Purpose: The aim of this study is to (1) demonstrate a graphical method and interpretation framework to extend performance evaluation beyond receiver operating characteristic curve analysis and (2) assess the impact of disease prevalence and variability in training and testing sets, particularly when a specific operating point is used.

Approach: The proposed performance metric curves (PMCs) simultaneously assess sensitivity, specificity, positive predictive value (PPV), and negative predictive value (NPV), and the 95% confidence intervals thereof, as a function of the threshold for the decision variable. We investigated the utility of PMCs using six example operating points associated with commonly used methods to select operating points (including the Youden index and maximum mutual information). As an example, we applied PMCs to the task of distinguishing between malignant and benign breast lesions using human-engineered radiomic features extracted from dynamic contrast-enhanced magnetic resonance images. The dataset had 1885 lesions, with the images acquired in 2015 and 2016 serving as the training set (1450 lesions) and those acquired in 2017 as the test set (435 lesions). Our study used this dataset in two ways: (1) the clinical dataset itself and (2) simulated datasets with features based on the clinical set but with five different disease prevalences. The median and 95% CI of the number of type I (false positive) and type II (false negative) errors were determined for each operating point of interest.

Results: PMCs from both the clinical and simulated datasets demonstrated that PMCs could support interpretation of the impact of decision threshold choice on type I and type II errors of classification, particularly relevant to prevalence.

Conclusion: PMCs allow simultaneous evaluation of the four performance metrics of sensitivity, specificity, PPV, and NPV as a function of the decision threshold. This may create a better understanding of two-class classifier performance in machine learning.

Purpose: Diagnosing breast cancer based on the distribution of calcifications is a visual task and thus prone to visual biases. We tested whether a recently discovered visual bias that has implications for breast cancer diagnosis would be present in expert radiologists, thereby validating the concern of this bias for accurate diagnoses.

Approach: We ran a vision experiment with expert radiologists and untrained observers to test the presence of visual bias when judging the spread of dots that resembled calcifications and when judging the spread of line orientations. We calculated visual bias scores for both groups for both tasks.

Results: Participants overestimated the spread of the dots and the spread of the line orientations. This bias, referred to as the variability overestimation effect, was of similar magnitudes in both expert radiologists and untrained observers. Even though the radiologists were better at both tasks, they were similarly biased compared with the untrained observers.

Conclusions: The results justify the concern of the variability overestimation effect for accurate diagnoses based on breast calcifications. Specifically, the bias is likely to lead to an increased number of false-negative results, thereby leading to delayed treatments.

Purpose: In recent years, the development and exploration of deeper and more complex deep learning models has been on the rise. However, the availability of large heterogeneous datasets to support efficient training of deep learning models is lacking. While linear image transformations for augmentation have been used traditionally, the recent development of generative adversarial networks (GANs) could theoretically allow us to generate an infinite amount of data from the real distribution to support deep learning model training. Recently, the Radiological Society of North America (RSNA) curated a multiclass hemorrhage detection challenge dataset that includes over 800,000 images for hemorrhage detection, but all high-performing models were trained using traditional data augmentation techniques. Given a wide variety of selections, the augmentation for image classification often follows a trial-and-error policy.

Approach: We designed conditional DCGAN (cDCGAN) and in parallel trained multiple popular GAN models to use as online augmentations and compared them to traditional augmentation methods for the hemorrhage case study.

Results: Our experimentations show that the super-minority, epidural hemorrhages with cDCGAN augmentation presented a minimum of 2  ×   improvement in their performance against the traditionally augmented model using the same classifier configuration.

Conclusion: This shows that for complex and imbalanced datasets, traditional data imbalancing solutions may not be sufficient and require more complex and diverse data augmentation methods such as GANs to solve.

Purpose: Multiparametric magnetic resonance imaging (mp-MRI) is being investigated for kidney cancer because of better soft tissue contrast ability. The necessity of manual labels makes the development of supervised kidney segmentation algorithms challenging for each mp-MRI protocol. Here, we developed a transfer learning-based approach to improve kidney segmentation on a small dataset of five other mp-MRI sequences.

Approach: We proposed a fully automated two-dimensional (2D) attention U-Net model for kidney segmentation on T1 weighted-nephrographic phase contrast enhanced (CE)-MRI (T1W-NG) dataset (N  =  108). The pretrained weights of T1W-NG kidney segmentation model transferred to five other distinct mp-MRI sequences model (T2W, T1W-in-phase (T1W-IP), T1W-out-of-phase (T1W-OP), T1W precontrast (T1W-PRE), and T1W-corticomedullary-CE (T1W-CM), N  =  50) and fine-tuned by unfreezing the layers. The individual model performances were evaluated with and without transfer-learning fivefold cross-validation on average Dice similarity coefficient (DSC), absolute volume difference, Hausdorff distance (HD), and center-of-mass distance (CD) between algorithm generated and manually segmented kidneys.

Results: The developed 2D attention U-Net model for T1W-NG produced kidney segmentation DSC of 89.34  ±  5.31  %  . Compared with randomly initialized weight models, the transfer learning-based models of five mp-MRI sequences showed average increase of 2.96% in DSC of kidney segmentation (p  =  0.001 to 0.006). Specifically, the transfer-learning approach increased average DSC on T2W from 87.19% to 89.90%, T1W-IP from 83.64% to 85.42%, T1W-OP from 79.35% to 83.66%, T1W-PRE from 82.05% to 85.94%, and T1W-CM from 85.65% to 87.64%.

Conclusions: We demonstrate that a pretrained model for automated kidney segmentation of one mp-MRI sequence improved automated kidney segmentation on five other additional sequences.

Purpose: In ultrasound imaging, interpolation is a key step in converting scanline data to brightness-mode (B-mode) images. Conventional methods, such as bilinear interpolation, do not fully capture the spatial dependence between data points, which leads to deviations from the underlying probability distribution at the interpolation points.

Approach: We propose Gaussian process (GP) regression as an improved method for ultrasound scanline interpolation. Using ultrasound scanlines acquired from two different ultrasound scanners during in vivo trials, we compare the scanline conversion accuracy of three standard interpolation methods with that of GP regression, measuring the peak signal-to-noise ratio (PSNR) and mean absolute error (MAE) for each method.

Results: The PSNR and MAE scores show that GP regression leads to more accurate scanline conversion compared to the nearest neighbor, bilinear, and cubic spline interpolation methods, for both datasets. Furthermore, limiting the interpolation window size of GP regression to 15 reduces computation time with minimal to no reduction in PSNR.

Conclusions: GP regression quantitatively leads to more accurate scanline conversion and provides uncertainty estimates at each of the interpolation points. Our windowing method reduces the computational cost of using GP regression for scanline conversion.