From code sharing to sharing of implementations: Advancing reproducible AI development for medical imaging through federated testing.

BACKGROUND: The reproducibility crisis in AI research remains a significant concern. While code sharing has been acknowledged as a step toward addressing this issue, our focus extends beyond this paradigm. In this work, we explore "federated testing" as an avenue for advancing reproducible AI research and development especially in medical imaging. Unlike federated learning, where a model is developed and refined on data from different centers, federated testing involves models developed by one team being deployed and evaluated by others, addressing reproducibility across various implementations. METHODS: Our study follows an exploratory design aimed at systematically evaluating the sources of discrepancies in shared model execution for medical imaging and outputs on the same input data, independent of generalizability analysis. We distributed the same model code to multiple independent centers, monitoring execution in different runtime environments while considering various real-world scenarios for pre- and post-processing steps. We analyzed deployment infrastructure by comparing the impact of different computational resources (GPU vs. CPU) on model performance. To assess federated testing in AI models for medical imaging, we performed a comparative evaluation across different centers, each with distinct pre- and post-processing steps and deployment environments, specifically targeting AI-driven positron emission tomography (PET) imaging segmentation. More specifically, we studied federated testing for an AI-based model for surrogate total metabolic tumor volume (sTMTV) segmentation in PET imaging: the AI algorithm, trained on maximum intensity projection (MIP) data, segments lymphoma regions and estimates sTMTV. RESULTS: Our study reveals that relying solely on open-source code sharing does not guarantee reproducible results due to variations in code execution, runtime environments, and incomplete input specifications. Deploying the segmentation model on local and virtual GPUs compared to using Docker containers showed no effect on reproducibility. However, significant sources of variability were found in data preparation and pre-/post- processing techniques for PET imaging. These findings underscore the limitations of code sharing alone in achieving consistent and accurate results in federated testing. CONCLUSION: Achieving consistently precise results in federated testing requires more than just sharing models through open-source code. Comprehensive pipeline sharing, including pre- and post-processing steps, is essential. Cloud-based platforms that automate these processes can streamline AI model testing across diverse locations. Standardizing protocols and sharing complete pipelines can significantly enhance the robustness and reproducibility of AI models.

Relating Macroscopic PET Radiomics Features to Microscopic Tumor Phenotypes Using a Stochastic Mathematical Model of Cellular Metabolism and Proliferation.

Cancers can manifest large variations in tumor phenotypes due to genetic and microenvironmental factors, which has motivated the development of quantitative radiomics-based image analysis with the aim to robustly classify tumor phenotypes in vivo. Positron emission tomography (PET) imaging can be particularly helpful in elucidating the metabolic profiles of tumors. However, the relatively low resolution, high noise, and limited PET data availability make it difficult to study the relationship between the microenvironment properties and metabolic tumor phenotype as seen on the images. Most of previously proposed digital PET phantoms of tumors are static, have an over-simplified morphology, and lack the link to cellular biology that ultimately governs the tumor evolution. In this work, we propose a novel method to investigate the relationship between microscopic tumor parameters and PET image characteristics based on the computational simulation of tumor growth. We use a hybrid, multiscale, stochastic mathematical model of cellular metabolism and proliferation to generate simulated cross-sections of tumors in vascularized normal tissue on a microscopic level. The generated longitudinal tumor growth sequences are converted to PET images with realistic resolution and noise. By changing the biological parameters of the model, such as the blood vessel density and conditions for necrosis, distinct tumor phenotypes can be obtained. The simulated cellular maps were compared to real histology slides of SiHa and WiDr xenografts imaged with Hoechst 33342 and pimonidazole. As an example application of the proposed method, we simulated six tumor phenotypes that contain various amounts of hypoxic and necrotic regions induced by a lack of oxygen and glucose, including phenotypes that are distinct on the microscopic level but visually similar in PET images. We computed 22 standardized Haralick texture features for each phenotype, and identified the features that could best discriminate the phenotypes with varying image noise levels. We demonstrated that "cluster shade" and "difference entropy" are the most effective and noise-resilient features for microscopic phenotype discrimination. Longitudinal analysis of the simulated tumor growth showed that radiomics analysis can be beneficial even in small lesions with a diameter of 3.5-4 resolution units, corresponding to 8.7-10.0 mm in modern PET scanners. Certain radiomics features were shown to change non-monotonically with tumor growth, which has implications for feature selection for tracking disease progression and therapy response.

Contralateral Hypertrophy Post Yttrium-90 Transarterial Radioembolization in Patients With Hepatocellular Carcinoma and Portal Vein Tumor Thrombus.

Objectives Contralateral hypertrophy of non-irradiated liver following Yttrium-90 (90Y) transarterial radioembolization (TARE) is increasingly recognized as an option to facilitate curative surgical resection in patients that would otherwise not be surgical candidates due to a small future liver remnant (FLR). This study aimed to investigate the correlation between patient features and liver hypertrophy and identify potential predictors for liver growth in patients with hepatocellular carcinoma (HCC) and portal vein tumor thrombus (PVTT) undergoing TARE. Methodology Twenty-three patients with HCC and PVTT were included. Contralateral liver hypertrophy was assessed at six months posttreatment based on CT or MRI imaging. Thirteen patient features were selected for statistical and prediction analysis. Univariate Spearman correlation and analysis of variance (ANOVA) tests were performed. Subsequently, four feature-selection methods based on multivariate analysis were used to improve model generalization performance. The selected features were applied to train linear regression models, with fivefold cross-validation to assess the performance of the predicted models. Results The ratio of disease-free target liver volume to spared liver volume and total liver volume showed the highest correlations with contralateral hypertrophy (P-values = 0.03 and 0.05, respectively). In three out of four feature-selection methods, the feature of disease-free target liver volume to total liver volume ratio was selected, having positive correlations with the outcome and suggesting that more hypertrophy may be expected when more volume of disease-free liver is irradiated. Conclusions Contralateral hypertrophy post-90Y TARE can be an option for facilitating surgical resection in patients with otherwise small FLR.

Multi-tracer PET correlation analysis reveals disease-specific patterns in Parkinson's disease and asymptomatic LRRK2 pathogenic variant carriers compared to healthy controls.

Several genetic pathogenic variants increase the risk of Parkinson's disease (PD) with pathogenic variants in the leucine-rich repeat kinase 2 (LRRK2) gene being among the most common. A joint pattern analysis based on multi-set canonical correlation analysis (MCCA) was utilized to extract PD and LRRK2 pathogenic variant-specific spatial patterns in relation to healthy controls (HCs) from multi-tracer Positron Emission Tomography (PET) data. Spatial patterns were extracted for individual subject cohorts, as well as for pooled subject cohorts, to explore whether complementary spatial patterns of dopaminergic denervation are different in the asymptomatic and symptomatic stages of PD. The MCCA results are also compared to the traditional univariate analysis, which serves as a reference. We identified PD-induced spatial distribution alterations common to DAT and VMAT2 in both asymptomatic LRRK2 pathogenic variant carriers and PD subjects. The inclusion of HCs in the analysis demonstrated that the dominant common PD-induced pattern is related to an overall dopaminergic terminal density denervation, followed by asymmetry and rostro-caudal gradient with deficits in the less affected side still being the best marker of disease progression. The analysis was able to capture a trend towards PD-related patterns in the LRRK2 pathogenic variant carrier cohort with increasing age in line with the known increased risk of this patient cohort to develop PD as they age. The advantage of this method thus resides in its ability to identify not only regional differences in tracer binding between groups, but also common disease-related alterations in the spatial distribution patterns of tracer binding, thus potentially capturing more complex aspects of disease induced alterations.

Evaluating Outcome Prediction via Baseline, End-of-Treatment, and Delta Radiomics on PET-CT Images of Primary Mediastinal Large B-Cell Lymphoma.

OBJECTIVES: Accurate outcome prediction is important for making informed clinical decisions in cancer treatment. In this study, we assessed the feasibility of using changes in radiomic features over time (Delta radiomics: absolute and relative) following chemotherapy, to predict relapse/progression and time to progression (TTP) of primary mediastinal large B-cell lymphoma (PMBCL) patients. MATERIAL AND METHODS: Given the lack of standard staging PET scans until 2011, only 31 out of 103 PMBCL patients in our retrospective study had both pre-treatment and end-of-treatment (EoT) scans. Consequently, our radiomics analysis focused on these 31 patients who underwent [18F]FDG PET-CT scans before and after R-CHOP chemotherapy. Expert manual lesion segmentation was conducted on their scans for delta radiomics analysis, along with an additional 19 EoT scans, totaling 50 segmented scans for single time point analysis. Radiomics features (on PET and CT), along with maximum and mean standardized uptake values (SUVmax and SUVmean), total metabolic tumor volume (TMTV), tumor dissemination (Dmax), total lesion glycolysis (TLG), and the area under the curve of cumulative standardized uptake value-volume histogram (AUC-CSH) were calculated. We additionally applied longitudinal analysis using radial mean intensity (RIM) changes. For prediction of relapse/progression, we utilized the individual coefficient approximation for risk estimation (ICARE) and machine learning (ML) techniques (K-Nearest Neighbor (KNN), Linear Discriminant Analysis (LDA), and Random Forest (RF)) including sequential feature selection (SFS) following correlation analysis for feature selection. For TTP, ICARE and CoxNet approaches were utilized. In all models, we used nested cross-validation (CV) (with 10 outer folds and 5 repetitions, along with 5 inner folds and 20 repetitions) after balancing the dataset using Synthetic Minority Oversampling TEchnique (SMOTE). RESULTS: To predict relapse/progression using Delta radiomics between the baseline (staging) and EoT scans, the best performances in terms of accuracy and F1 score (F1 score is the harmonic mean of precision and recall, where precision is the ratio of true positives to the sum of true positives and false positives, and recall is the ratio of true positives to the sum of true positives and false negatives) were achieved with ICARE (accuracy = 0.81 ± 0.15, F1 = 0.77 ± 0.18), RF (accuracy = 0.89 ± 0.04, F1 = 0.87 ± 0.04), and LDA (accuracy = 0.89 ± 0.03, F1 = 0.89 ± 0.03), that are higher compared to the predictive power achieved by using only EoT radiomics features. For the second category of our analysis, TTP prediction, the best performer was CoxNet (LASSO feature selection) with c-index = 0.67 ± 0.06 when using baseline + Delta features (inclusion of both baseline and Delta features). The TTP results via Delta radiomics were comparable to the use of radiomics features extracted from EoT scans for TTP analysis (c-index = 0.68 ± 0.09) using CoxNet (with SFS). The performance of Deauville Score (DS) for TTP was c-index = 0.66 ± 0.09 for n = 50 and 0.67 ± 03 for n = 31 cases when using EoT scans with no significant differences compared to the radiomics signature from either EoT scans or baseline + Delta features (p-value> 0.05). CONCLUSION: This work demonstrates the potential of Delta radiomics and the importance of using EoT scans to predict progression and TTP from PMBCL [18F]FDG PET-CT scans.

TMTV-Net: fully automated total metabolic tumor volume segmentation in lymphoma PET/CT images - a multi-center generalizability analysis.

PURPOSE: Total metabolic tumor volume (TMTV) segmentation has significant value enabling quantitative imaging biomarkers for lymphoma management. In this work, we tackle the challenging task of automated tumor delineation in lymphoma from PET/CT scans using a cascaded approach. METHODS: Our study included 1418 2-[18F]FDG PET/CT scans from four different centers. The dataset was divided into 900 scans for development/validation/testing phases and 518 for multi-center external testing. The former consisted of 450 lymphoma, lung cancer, and melanoma scans, along with 450 negative scans, while the latter consisted of lymphoma patients from different centers with diffuse large B cell, primary mediastinal large B cell, and classic Hodgkin lymphoma cases. Our approach involves resampling PET/CT images into different voxel sizes in the first step, followed by training multi-resolution 3D U-Nets on each resampled dataset using a fivefold cross-validation scheme. The models trained on different data splits were ensemble. After applying soft voting to the predicted masks, in the second step, we input the probability-averaged predictions, along with the input imaging data, into another 3D U-Net. Models were trained with semi-supervised loss. We additionally considered the effectiveness of using test time augmentation (TTA) to improve the segmentation performance after training. In addition to quantitative analysis including Dice score (DSC) and TMTV comparisons, the qualitative evaluation was also conducted by nuclear medicine physicians. RESULTS: Our cascaded soft-voting guided approach resulted in performance with an average DSC of 0.68 ± 0.12 for the internal test data from developmental dataset, and an average DSC of 0.66 ± 0.18 on the multi-site external data (n = 518), significantly outperforming (p < 0.001) state-of-the-art (SOTA) approaches including nnU-Net and SWIN UNETR. While TTA yielded enhanced performance gains for some of the comparator methods, its impact on our cascaded approach was found to be negligible (DSC: 0.66 ± 0.16). Our approach reliably quantified TMTV, with a correlation of 0.89 with the ground truth (p < 0.001). Furthermore, in terms of visual assessment, concordance between quantitative evaluations and clinician feedback was observed in the majority of cases. The average relative error (ARE) and the absolute error (AE) in TMTV prediction on external multi-centric dataset were ARE = 0.43 ± 0.54 and AE = 157.32 ± 378.12 (mL) for all the external test data (n = 518), and ARE = 0.30 ± 0.22 and AE = 82.05 ± 99.78 (mL) when the 10% outliers (n = 53) were excluded. CONCLUSION: TMTV-Net demonstrates strong performance and generalizability in TMTV segmentation across multi-site external datasets, encompassing various lymphoma subtypes. A negligible reduction of 2% in overall performance during testing on external data highlights robust model generalizability across different centers and cancer types, likely attributable to its training with resampled inputs. Our model is publicly available, allowing easy multi-site evaluation and generalizability analysis on datasets from different institutions.

PSMA-Hornet: Fully-automated, multi-target segmentation of healthy organs in PSMA PET/CT images.

BACKGROUND: Prostate-specific membrane antigen (PSMA) PET imaging represents a valuable source of information reflecting disease stage, response rate, and treatment optimization options, particularly with PSMA radioligand therapy. Quantification of radiopharmaceutical uptake in healthy organs from PSMA images has the potential to minimize toxicity by extrapolation of the radiation dose delivery towards personalization of therapy. However, segmentation and quantification of uptake in organs requires labor-intensive organ delineations that are often not feasible in the clinic nor scalable for large clinical trials. PURPOSE: In this work we develop and test the PSMA Healthy organ segmentation network (PSMA-Hornet), a fully-automated deep neural net for simultaneous segmentation of 14 healthy organs representing the normal biodistribution of [18 F]DCFPyL on PET/CT images. We also propose a modified U-net architecture, a self-supervised pre-training method for PET/CT images, a multi-target Dice loss, and multi-target batch balancing to effectively train PSMA-Hornet and similar networks. METHODS: The study used manually-segmented [18 F]DCFPyL PET/CT images from 100 subjects, and 526 similar images without segmentations. The unsegmented images were used for self-supervised model pretraining. For supervised training, Monte-Carlo cross-validation was used to evaluate the network performance, with 85 subjects in each trial reserved for model training, 5 for validation, and 10 for testing. Image segmentation and quantification metrics were evaluated on the test folds with respect to manual segmentations by a nuclear medicine physician, and compared to inter-rater agreement. The model's segmentation performance was also evaluated on a separate set of 19 images with high tumor load. RESULTS: With our best model, the lowest mean Dice coefficient on the test set was 0.826 for the sublingual gland, and the highest was 0.964 for liver. The highest mean error in tracer uptake quantification was 13.9% in the sublingual gland. Self-supervised pretraining improved training convergence, train-to-test generalization, and segmentation quality. In addition, we found that a multi-target network produced significantly higher segmentation accuracy than single-organ networks. CONCLUSIONS: The developed network can be used to automatically obtain high-quality organ segmentations for PSMA image analysis tasks. It can be used to reproducibly extract imaging data, and holds promise for clinical applications such as personalized radiation dosimetry and improved radioligand therapy.

Automatic segmentation of prostate cancer metastases in PSMA PET/CT images using deep neural networks with weighted batch-wise dice loss.

PURPOSE: Automatic and accurate segmentation of lesions in images of metastatic castration-resistant prostate cancer has the potential to enable personalized radiopharmaceutical therapy and advanced treatment response monitoring. The aim of this study is to develop a convolutional neural networks-based framework for fully-automated detection and segmentation of metastatic prostate cancer lesions in whole-body PET/CT images. METHODS: 525 whole-body PET/CT images of patients with metastatic prostate cancer were available for the study, acquired with the [18F]DCFPyL radiotracer that targets prostate-specific membrane antigen (PSMA). U-Net (1)-based convolutional neural networks (CNNs) were trained to identify lesions on paired axial PET/CT slices. Baseline models were trained using batch-wise dice loss, as well as the proposed weighted batch-wise dice loss (wDice), and the lesion detection performance was quantified, with a particular emphasis on lesion size, intensity, and location. We used 418 images for model training, 30 for model validation, and 77 for model testing. In addition, we allowed our model to take n = 0,2, …, 12 neighboring axial slices to examine how incorporating greater amounts of 3D context influences model performance. We selected the optimal number of neighboring axial slices that maximized the detection rate on the 30 validation images, and trained five neural networks with different architectures. RESULTS: Model performance was evaluated using the detection rate, Dice similarity coefficient (DSC) and sensitivity. We found that the proposed wDice loss significantly improved the lesion detection rate, lesion-wise DSC and lesion-wise sensitivity compared to the baseline, with corresponding average increases of 0.07 (p-value = 0.01), 0.03 (p-value = 0.01) and 0.04 (p-value = 0.01), respectively. The inclusion of the first two neighboring axial slices in the input likewise increased the detection rate by 0.17, lesion-wise DSC by 0.05, and lesion-wise mean sensitivity by 0.16. However, there was a minimal effect from including more distant neighboring slices. We ultimately chose to use a number of neighboring slices equal to 2 and the wDice loss function to train our final model. To evaluate the model's performance, we trained three models using identical hyperparameters on three different data splits. The results showed that, on average, the model was able to detect 80% of all testing lesions, with a detection rate of 93% for lesions with maximum standardized uptake values (SUVmax) greater than 5.0. In addition, the average median lesion-wise DSC was 0.51 and 0.60 for all the lesions and lesions with SUVmax>5.0, respectively, on the testing set. Four additional neural networks with different architectures were trained, and they both yielded stronger performance of segmenting lesions whose SUVmax>5.0 compared to the rest of lesions. CONCLUSION: Our results demonstrate that prostate cancer metastases in PSMA PET/CT images can be detected and segmented using CNNs. The segmentation performance strongly depends on the intensity, size, and the location of lesions, and can be improved by using specialized loss functions. Specifically, the models performed best in detection of lesions with SUVmax>5.0. Another challenge was to accurately segment lesions close to the bladder. Future work will focus on improving the detection of lesions with lower SUV values by designing custom loss functions that take into account the lesion intensity, using additional data augmentation techniques, and reducing the number of false lesions by developing methods to better separate signal from noise.

Spatiotemporal patterns of putaminal dopamine processing in Parkinson's disease: A multi-tracer positron emission tomography study.

Alterations in different aspects of dopamine processing may exhibit different progressive behaviours throughout the course of Parkinson's disease. We used a novel data-driven multivariate approach to quantify and compare spatiotemporal patterns related to different aspects of dopamine processing from cross-sectional Parkinson's subjects obtained with: 1) 69 [11C]±dihydrotetrabenazine (DTBZ) scans, most closely related to dopaminergic denervation; 2) 73 [11C]d-threo-methylphenidate (MP) scans, marker of dopamine transporter density; 3) 50 6-[18F]fluoro-l-DOPA (FD) scans, marker of dopamine synthesis and storage. The anterior-posterior gradient in the putamen was identified as the most salient feature associated with disease progression, however the temporal progression of the spatial gradient was different for the three tracers. The expression of the anterior-posterior gradient was the highest for FD at disease onset compared to that of DTBZ and MP (P = 0.018 and P = 0.047 respectively), but decreased faster (P = 0.006) compared to that of DTBZ. The gradient expression for MP was initially similar but decreased faster (P = 0.015) compared to that for DTBZ. These results reflected unique temporal behaviours of regulatory mechanisms related to dopamine synthesis (FD) and reuptake (MP). While the relative early disease upregulation of dopamine synthesis in the anterior putamen prevalent likely extends to approximately 10 years after symptom onset, the presumed downregulation of dopamine transporter density may play a compensatory role in the prodromal/earliest disease stages only.

Testing the Ability of Convolutional Neural Networks to Learn Radiomic Features.

BACKGROUND AND OBJECTIVE: Radiomics and deep learning have emerged as two distinct approaches to medical image analysis. However, their relative expressive power remains largely unknown. Theoretically, hand-crafted radiomic features represent a mere subset of features that neural networks can approximate, thus making deep learning a more powerful approach. On the other hand, automated learning of hand-crafted features may require a prohibitively large number of training samples. Here we directly test the ability of convolutional neural networks (CNNs) to learn and predict the intensity, shape, and texture properties of tumors as defined by standardized radiomic features. METHODS: Conventional 2D and 3D CNN architectures with an increasing number of convolutional layers were trained to predict the values of 16 standardized radiomic features from real and synthetic PET images of tumors, and tested. In addition, several ImageNet-pretrained advanced networks were tested. A total of 4000 images were used for training, 500 for validation, and 500 for testing. RESULTS: Features quantifying size and intensity were predicted with high accuracy, while shape irregularity and heterogeneity features had very high prediction errors and generalized poorly. For example, mean normalized prediction error of tumor diameter with a 5-layer CNN was 4.23 ± 0.25, while the error for tumor sphericity was 15.64 ± 0.93. We additionally found that learning shape features required an order of magnitude more samples compared to intensity and size features. CONCLUSIONS: Our findings imply that CNNs trained to perform various image-based clinical tasks may generally under-utilize the shape and texture information that is more easily captured by radiomics. We speculate that to improve the CNN performance, shape and texture features can be computed explicitly and added as auxiliary variables to the networks, or supplied as synthetic inputs.

Quantitative evaluation of PSMA PET imaging using a realistic anthropomorphic phantom and shell-less radioactive epoxy lesions.

BACKGROUND: Positron emission tomography (PET) with prostate specific membrane antigen (PSMA) have shown superior performance in detecting metastatic prostate cancers. Relative to [18F]fluorodeoxyglucose ([18F]FDG) PET images, PSMA PET images tend to visualize significantly higher-contrast focal lesions. We aim to evaluate segmentation and reconstruction algorithms in this emerging context. Specifically, Bayesian or maximum a posteriori (MAP) image reconstruction, compared to standard ordered subsets expectation maximization (OSEM) reconstruction, has received significant interest for its potential to reach convergence with minimal noise amplifications. However, few phantom studies have evaluated the quantitative accuracy of such reconstructions for high contrast, small lesions (sub-10 mm) that are typically observed in PSMA images. In this study, we cast 3 mm-16-mm spheres using epoxy resin infused with a long half-life positron emitter (sodium-22; 22Na) to simulate prostate cancer metastasis. The anthropomorphic Probe-IQ phantom, which features a liver, bladder, lungs, and ureters, was used to model relevant anatomy. Dynamic PET acquisitions were acquired and images were reconstructed with OSEM (varying subsets and iterations) and BSREM (varying ß parameters), and the effects on lesion quantitation were evaluated. RESULTS: The 22Na lesions were scanned against an aqueous solution containing fluorine-18 (18F) as the background. Regions-of-interest were drawn with MIM Software using 40% fixed threshold (40% FT) and a gradient segmentation algorithm (MIM's PET Edge+). Recovery coefficients (RCs) (max, mean, peak, and newly defined "apex"), metabolic tumour volume (MTV), and total tumour uptake (TTU) were calculated for each sphere. SUVpeak and SUVapex had the most consistent RCs for different lesion-to-background ratios and reconstruction parameters. The gradient-based segmentation algorithm was more accurate than 40% FT for determining MTV and TTU, particularly for lesions [Formula: see text] 6 mm in diameter (R2 = 0.979-0.996 vs. R2 = 0.115-0.527, respectively). CONCLUSION: An anthropomorphic phantom was used to evaluate quantitation for PSMA PET imaging of metastatic prostate cancer lesions. BSREM with ß = 200-400 and OSEM with 2-5 iterations resulted in the most accurate and robust measurements of SUVmean, MTV, and TTU for imaging conditions in 18F-PSMA PET/CT images. SUVapex, a hybrid metric of SUVmax and SUVpeak, was proposed for robust, accurate, and segmentation-free quantitation of lesions for PSMA PET.

Clinical Application of Artificial Intelligence in Positron Emission Tomography: Imaging of Prostate Cancer.

PET imaging with targeted novel tracers has been commonly used in the clinical management of prostate cancer. The use of artificial intelligence (AI) in PET imaging is a relatively new approach and in this review article, we will review the current trends and categorize the currently available research into the quantification of tumor burden within the organ, evaluation of metastatic disease, and translational/supplemental research which aims to improve other AI research efforts.

Radiomics in PET Imaging:: A Practical Guide for Newcomers.

Radiomics has undergone considerable development in recent years. In PET imaging, very promising results concerning the ability of handcrafted features to predict the biological characteristics of lesions and to assess patient prognosis or response to treatment have been reported in the literature. This article presents a checklist for designing a reliable radiomic study, gives an overview of the steps of the pipeline, and outlines approaches for data harmonization. Tips are provided for critical reading of the content of articles. The advantages and limitations of handcrafted radiomics compared with deep-learning approaches for the characterization of PET images are also discussed.

Design of an anthropomorphic PET phantom with elastic lungs and respiration modeling.

PURPOSE: Respiratory motion during positron emission tomography (PET) scans can be a major detriment to image quality in oncological imaging. The impact of motion on lesion quantification and detectability can be assessed using phantoms with realistic anatomy representation and motion modeling. In this work, we develop an anthropomorphic phantom for PET imaging that combines anatomic fidelity and a realistic breathing mechanism with deformable lungs. METHODS: We start from a previously developed anatomically accurate but static phantom of a human torso, and add elastic lungs with a highly controllable actuation mechanism which replicates the physics of breathing. The space outside the lungs is filled with a radioactive water solution. To maintain anatomical accuracy and realistic gamma ray attenuation in the torso, all motion mechanisms and actuators are positioned outside of the phantom compartment. The actuation mechanism can produce custom respiratory waveforms with breathing rates up to 25 breaths per minute and tidal volumes up to 1200 mL. RESULTS: Several tests were performed to validate the performance of the phantom assembly, in which the phantom was filled with water and given respiratory waveforms to execute. All parts demonstrated expected performance. Force requirements were not exceeded and no leaks were detected, although continued use of the phantom is required to evaluate wear. The motion of the lungs was determined to be within a reasonable realistic range. CONCLUSIONS: The full mechanical design is described in this paper, as well as a software application with graphical user interface which was developed to plan and visualize respiratory patterns. Both are available online as open source files. The developed phantom will facilitate future work in evaluating the impact of respiratory motion on lesion quantification and detectability in clinical practice.

Dynamic PET image reconstruction utilizing intrinsic data-driven HYPR4D denoising kernel.

PURPOSE: Reconstructed PET images are typically noisy, especially in dynamic imaging where the acquired data are divided into several short temporal frames. High noise in the reconstructed images translates to poor precision/reproducibility of image features. One important role of "denoising" is therefore to improve the precision of image features. However, typical denoising methods achieve noise reduction at the expense of accuracy. In this work, we present a novel four-dimensional (4D) denoised image reconstruction framework, which we validate using 4D simulations, experimental phantom, and clinical patient data, to achieve 4D noise reduction while preserving spatiotemporal patterns/minimizing error introduced by denoising. METHODS: Our proposed 4D denoising operator/kernel is based on HighlY constrained backPRojection (HYPR), which is applied either after each update of OSEM reconstruction of dynamic 4D PET data or within the recently proposed kernelized reconstruction framework inspired by kernel methods in machine learning. Our HYPR4D kernel makes use of the spatiotemporal high frequency features extracted from a 4D composite, generated within the reconstruction, to preserve the spatiotemporal patterns and constrain the 4D noise increment of the image estimate. RESULTS: Results from simulations, experimental phantom, and patient data showed that the HYPR4D kernel with our proposed 4D composite outperformed other denoising methods, such as the standard OSEM with spatial filter, OSEM with 4D filter, and HYPR kernel method with the conventional 3D composite in conjunction with recently proposed High Temporal Resolution kernel (HYPRC3D-HTR), in terms of 4D noise reduction while preserving the spatiotemporal patterns or 4D resolution within the 4D image estimate. Consequently, the error in outcome measures obtained from the HYPR4D method was less dependent on the region size, contrast, and uniformity/functional patterns within the target structures compared to the other methods. For outcome measures that depend on spatiotemporal tracer uptake patterns such as the nondisplaceable Binding Potential (BPND ), the root mean squared error in regional mean of voxel BPND values was reduced from ~8% (OSEM with spatial or 4D filter) to ~3% using HYPRC3D-HTR and was further reduced to ~2% using our proposed HYPR4D method for relatively small target structures (~10 mm in diameter). At the voxel level, HYPR4D produced two to four times lower mean absolute error in BPND relative to HYPRC3D-HTR. CONCLUSION: As compared to conventional methods, our proposed HYPR4D method can produce more robust and accurate image features without requiring any prior information.

A Monte Carlo approach for improving transient dopamine release detection sensitivity.

Current methods using a single PET scan to detect voxel-level transient dopamine release-using F-test (significance) and cluster size thresholding-have limited detection sensitivity for clusters of release small in size and/or having low release levels. Specifically, simulations show that voxels with release near the peripheries of such clusters are often rejected-becoming false negatives and ultimately distorting the F-distribution of rejected voxels. We suggest a Monte Carlo method that incorporates these two observations into a cost function, allowing erroneously rejected voxels to be accepted under specified criteria. In simulations, the proposed method improves detection sensitivity by up to 50% while preserving the cluster size threshold, or up to 180% when optimizing for sensitivity. A further parametric-based voxelwise thresholding is then suggested to better estimate the release dynamics in detected clusters. We apply the Monte Carlo method to a pilot scan from a human gambling study, where additional parametrically unique clusters are detected as compared to the current best methods-results consistent with our simulations.

Detection of transient neurotransmitter response using personalized neural networks.

Measurement of stimulus-induced dopamine release and other types of transient neurotransmitter response (TNR) from dynamic positron emission tomography (PET) images typically suffers from limited detection sensitivity and high false positive (FP) rates. Measurement of TNR of a voxel-level can be particularly problematic due to high image noise. In this work, we perform voxel-level TNR detection using artificial neural networks (ANN) and compare their performance to previously used standard statistical tests. Different ANN architectures were trained and tested using simulated and real human PET imaging data, obtained with the tracer [11C]raclopride (a D2 receptor antagonist). A distinguishing feature of our approach is the use of 'personalized' ANNs that are designed to operate on the image from a specific subject and scan. Training of personalized ANNs was performed using simulated images that have been matched with the acquired image in terms of the signal, resolution, and noise. In our tests of TNR detection performance, the F-test of the linear parametric neurotransmitter PET model fit residuals was used as the reference method. For a moderate TNR magnitude, the areas under the receiver operating characteristic curves in simulated tests were 0.64 for the F-test and 0.77-0.79 for the best ANNs. At a fixed FP rate of 0.01, the true positive rates were 0.6 for the F-test and 0.8-0.9 for the ANNs. The F-test detected on average 28% of a 8.4 mm cluster with a strong TNR, while the best ANN detected 47%. When applied to a real image, no significant abnormalities in the ANN outputs were observed. These results demonstrate that personalized ANNs may offer a greater detection sensitivity of dopamine release and other types of TNR compared to previously used method based on the F-test.

Novel data-driven, equation-free method captures spatio-temporal patterns of neurodegeneration in Parkinson's disease: Application of dynamic mode decomposition to PET.

Most neurodegenerative disorders are characterized by progressive loss of neurons throughout the course of disease in the form of specific spatio-temporal patterns. To capture and quantify these coherent patterns across both space and time, traditionally one would either fit a pre-defined model with spatial and temporal parameters or apply analysis in the spatial and temporal domains separately. In this work, we introduce and validate the use of dynamic mode decomposition (DMD), a data-driven multivariate approach, to extract coupled spatio-temporal patterns simultaneously. We apply the method to examine progressive dopaminergic degeneration in 41 patients with Parkinson's disease (PD) using [11C](±)dihydrotetrabenazine (DTBZ) Positron Emission Tomography (PET). DMD decomposed the progressive dopaminergic changes in the putamen into two orthogonal temporal progression curves associated with distinct spatial patterns: 1) an anterior-posterior gradient, the expression of which decreased gradually with disease progression with a higher initial expression in the less affected side; 2) a dorsal-ventral gradient in the less affected side, which was present in early disease stage only. In the caudate, we found a head-tail gradient analogous to the anterior-posterior gradient seen in the putamen; as in the putamen, the expression of this gradient decreased gradually with disease progression with higher expression in the less affected side. Our results with DTBZ PET data show the applicability and relevance of the proposed method for extracting spatio-temporal patterns of neurotransmitter changes due to neurodegeneration. The method is able to decompose known PD-induced dopaminergic denervation into orthogonal (and thus loosely independent) temporal curves, which may be able to reflect and separate either different mechanisms underlying disease progression and disease initiation, or differential involvement of striatal sub-regions at different disease stages, in a completely data driven way. It is expected that this method can be easily extended to other PET tracers and neurodegenerative disorders and may help to elucidate disease mechanisms in more details compared to traditional approaches.

Machine learning methods for optimal prediction of motor outcome in Parkinson's disease.

PURPOSE: It is vital to appropriately power clinical trials towards discovery of novel disease-modifying therapies for Parkinson's disease (PD). Thus, it is critical to improve prediction of outcome in PD patients. METHODS: We systematically probed a range of robust predictor algorithms, aiming to find best combinations of features for significantly improved prediction of motor outcome (MDS-UPDRS-III) in PD. We analyzed 204 PD patients with 18 features (clinical measures; dopamine-transporter (DAT) SPECT imaging measures), performing different randomized arrangements and utilizing data from 64%/6%/30% of patients in each arrangement for training/training validation/final testing. We pursued 3 approaches: i) 10 predictor algorithms (accompanied with automated machine learning hyperparameter tuning) were first applied on 32 experimentally created combinations of 18 features, ii) we utilized Feature Subset Selector Algorithms (FSSAs) for more systematic initial feature selection, and iii) considered all possible combinations between 18 features (262,143 states) to assess contributions of individual features. RESULTS: A specific set (set 18) applied to the LOLIMOT (Local Linear Model Trees) predictor machine resulted in the lowest absolute error 4.32 ± 0.19, when we firstly experimentally created 32 combinations of 18 features. Subsequently, 2 FSSAs (Genetic Algorithm (GA) and Ant Colony Optimization (ACO)) selecting 5 features, combined with LOLIMOT, reached an error of 4.15 ± 0.46. Our final analysis indicated that longitudinal motor measures (MDS-UPDRS-III years 0 and 1) were highly significant predictors of motor outcome. CONCLUSIONS: We demonstrate excellent prediction of motor outcome in PD patients by employing automated hyperparameter tuning and optimal utilization of FSSAs and predictor algorithms.

Use of a Tracer-Specific Deep Artificial Neural Net to Denoise Dynamic PET Images.

Application of kinetic modeling (KM) on a voxel level in dynamic PET images frequently suffers from high levels of noise, drastically reducing the precision of parametric image analysis. In this paper, we investigate the use of machine learning and artificial neural networks to denoise dynamic PET images. We train a deep denoising autoencoder (DAE) using noisy and noise-free spatiotemporal image patches, extracted from the simulated images of [11C]raclopride, a dopamine D2 receptor agonist. The DAE-processed dynamic and corresponding parametric images (simulated and acquired) are compared with those obtained with conventional denoising techniques, including temporal and spatial Gaussian smoothing, iterative spatiotemporal smoothing/deconvolution, and the highly constrained backprojection processing (HYPR). The simulated (acquired) parametric image non-uniformity was 7.75% (19.49%) with temporal and 5.90% (14.50%) with spatial smoothing, 5.82% (16.21%) with smoothing/deconvolution, 5.49% (13.38%) with HYPR, and 3.52% (11.41%) with DAE. The DAE also produced the best results in terms of the coefficient of variation of voxel values and structural similarity index. Denoising-induced bias in the regional mean binding potential was 7.8% with temporal and 26.31% with spatial smoothing, 28.61% with smoothing/deconvolution, 27.63% with HYPR, and 14.8% with DAE. When the test data did not match the training data, erroneous outcomes were obtained. Our results demonstrate that a deep DAE can provide a substantial reduction in the voxel-level noise compared with the conventional spatiotemporal denoising methods while introducing a similar or lower amount of bias. The better DAE performance comes at the cost of lower generality and requiring appropriate training data.