SpatialSPM: statistical parametric mapping for the comparison of gene expression pattern images in multiple spatial transcriptomic datasets.

Spatial transcriptomic (ST) techniques help us understand the gene expression levels in specific parts of tissues and organs, providing insights into their biological functions. Even though ST dataset provides information on the gene expression and its location for each sample, it is challenging to compare spatial gene expression patterns across tissue samples with different shapes and coordinates. Here, we propose a method, SpatialSPM, that reconstructs ST data into multi-dimensional image matrices to ensure comparability across different samples through spatial registration process. We demonstrated the applicability of this method by kidney and mouse olfactory bulb datasets as well as mouse brain ST datasets to investigate and directly compare gene expression in a specific anatomical region of interest, pixel by pixel, across various biological statuses. Beyond traditional analyses, SpatialSPM is capable of generating statistical parametric maps, including T-scores and Pearson correlation coefficients. This feature enables the identification of specific regions exhibiting differentially expressed genes across tissue samples, enhancing the depth and specificity of ST studies. Our approach provides an efficient way to analyze ST datasets and may offer detailed insights into various biological conditions.

Spatial transcriptomic brain imaging reveals the effects of immunomodulation therapy on specific regional brain cells in a mouse dementia model.

Increasing evidence of brain-immune crosstalk raises expectations for the efficacy of novel immunotherapies in Alzheimer's disease (AD), but the lack of methods to examine brain tissues makes it difficult to evaluate therapeutics. Here, we investigated the changes in spatial transcriptomic signatures and brain cell types using the 10x Genomics Visium platform in immune-modulated AD models after various treatments. To proceed with an analysis suitable for barcode-based spatial transcriptomics, we first organized a workflow for segmentation of neuroanatomical regions, establishment of appropriate gene combinations, and comprehensive review of altered brain cell signatures. Ultimately, we investigated spatial transcriptomic changes following administration of immunomodulators, NK cell supplements and an anti-CD4 antibody, which ameliorated behavior impairment, and designated brain cells and regions showing probable associations with behavior changes. We provided the customized analytic pipeline into an application named STquantool. Thus, we anticipate that our approach can help researchers interpret the real action of drug candidates by simultaneously investigating the dynamics of all transcripts for the development of novel AD therapeutics.

Distinct subtypes of spatial brain metabolism patterns in Alzheimer's disease identified by deep learning-based FDG PET clusters.

PURPOSE: Alzheimer's disease (AD) is a heterogeneous disease that presents a broad spectrum of clinicopathologic profiles. To date, objective subtyping of AD independent of disease progression using brain imaging has been required. Our study aimed to extract representations of unique brain metabolism patterns different from disease progression to identify objective subtypes of AD. METHODS: A total of 3620 FDG brain PET images with AD, mild cognitive impairment (MCI), and cognitively normal (CN) were obtained from the ADNI database from 1607 participants at enrollment and follow-up visits. A conditional variational autoencoder model was trained on FDG brain PET images of AD patients with the corresponding condition of AD severity score. The k-means algorithm was applied to generate clusters from the encoded representations. The trained deep learning-based cluster model was also transferred to FDG PET of MCI patients and predicted the prognosis of subtypes for conversion from MCI to AD. Spatial metabolism patterns, clinical and biological characteristics, and conversion rate from MCI to AD were compared across the subtypes. RESULTS: Four distinct subtypes of spatial metabolism patterns in AD with different brain pathologies and clinical profiles were identified: (i) angular, (ii) occipital, (iii) orbitofrontal, and (iv) minimal hypometabolic patterns. The deep learning model was also successfully transferred for subtyping MCI, and significant differences in frequency (P < 0.001) and risk of conversion (log-rank P < 0.0001) from MCI to AD were observed across the subtypes, highest in S2 (35.7%) followed by S1 (23.4%). CONCLUSION: We identified distinct subtypes of AD with different clinicopathologic features. The deep learning-based approach to distinguish AD subtypes on FDG PET could have implications for predicting individual outcomes and provide a clue to understanding the heterogeneous pathophysiology of AD.

Preoperative evaluation of mediastinal lymph nodes in non-small cell lung cancer using [68Ga]FAPI-46 PET/CT: a prospective pilot study.

PURPOSE: Mediastinal nodal staging is crucial for surgical candidate selection in non-small cell lung cancer (NSCLC), but conventional imaging has limitations often necessitating invasive staging. We investigated the additive clinical value of fibroblast activation protein inhibitor (FAPI) PET/CT, an imaging technique targeting fibroblast activation protein, for mediastinal nodal staging of NSCLC. METHODS: In this prospective pilot study, we enrolled patients scheduled for surgical resection of NSCLC based on specific criteria designed to align with indications for invasive staging procedures. Patients were included when meeting at least one of the following: (1) presence of FDG-positive N2 lymph nodes, (2) clinical N1 stage, (3) central tumor location or tumor diameter of ≥ 3 cm, and (4) adenocarcinoma exhibiting high FDG uptake. [68Ga]FAPI-46 PET/CT was performed before surgery after a staging workup including [18F]FDG PET/CT. The diagnostic accuracy of [68Ga]FAPI-46 PET/CT for "N2" nodes was assessed through per-patient visual assessment and per-station quantitative analysis using histopathologic results as reference standards. RESULTS: Twenty-three patients with 75 nodal stations were analyzed. Histopathologic examination confirmed that nine patients (39.1%) were N2-positive. In per-patient assessment, [68Ga]FAPI-46 PET/CT successfully identified metastasis in eight patients (sensitivity 0.89 (0.52-1.00)), upstaging three patients compared to [18F]FDG PET/CT. [18F]FDG PET/CT detected FDG-avid nodes in six (42.8%) of 14 N2-negative patients. Among them, five were considered non-metastatic based on calcification and distribution pattern, and one was considered metastatic. In contrast, [68Ga]FAPI-46 PET/CT correctly identified all non-metastatic patients solely based on tracer avidity. In per-station analysis, [68Ga]FAPI-46 PET/CT discriminated metastasis more effectively compared to [18F]FDG PET/CT-based (AUC of ROC curve 0.96 (0.88-0.99) vs. 0.68 (0.56-0.78), P < 0.001). CONCLUSION: [68Ga]FAPI-46 PET/CT holds promise as an imaging tool for preoperative mediastinal nodal staging in NSCLC, with improved sensitivity and the potential to reduce false-positive results, optimizing the need for invasive staging procedures.

CellDART: cell type inference by domain adaptation of single-cell and spatial transcriptomic data.

Deciphering the cellular composition in genome-wide spatially resolved transcriptomic data is a critical task to clarify the spatial context of cells in a tissue. In this study, we developed a method, CellDART, which estimates the spatial distribution of cells defined by single-cell level data using domain adaptation of neural networks and applied it to the spatial mapping of human lung tissue. The neural network that predicts the cell proportion in a pseudospot, a virtual mixture of cells from single-cell data, is translated to decompose the cell types in each spatial barcoded region. First, CellDART was applied to a mouse brain and a human dorsolateral prefrontal cortex tissue to identify cell types with a layer-specific spatial distribution. Overall, the proposed approach showed more stable and higher accuracy with short execution time compared to other computational methods to predict the spatial location of excitatory neurons. CellDART was capable of decomposing cellular proportion in mouse hippocampus Slide-seq data. Furthermore, CellDART elucidated the cell type predominance defined by the human lung cell atlas across the lung tissue compartments and it corresponded to the known prevalent cell types. CellDART is expected to help to elucidate the spatial heterogeneity of cells and their close interactions in various tissues.

99mTc-MAA accumulation within tumor in preoperative lung perfusion SPECT/CT associated with occult lymph node metastasis in patients with clinically N0 non-small cell lung cancer.

BACKGROUND: 99mTc-MAA accumulation within the tumor representing pulmonary arterial perfusion, which is variable and may have a clinical significance. We evaluated the prognostic significance of 99mTc-MAA distribution within the tumor in non-small cell lung cancer (NSCLC) patients in terms of detecting occult nodal metastasis and lymphovascular invasion, as well as predicting the recurrence-free survival (RFS). METHODS: Two hundred thirty-nine NSCLC patients with clinical N0 status who underwent preoperative lung perfusion SPECT/CT were retrospectively evaluated and classified according to the visual grading of 99mTc-MAA accumulation in the tumor. Visual grade was compared with the quantitative parameter, standardized tumor to lung ratio (TLR). The predictive value of 99mTc-MAA accumulation with occult nodal metastasis, lymphovascular invasion, and RFS was assessed. RESULTS: Eighty-nine (37.2%) patients showed 99mTc-MAA accumulation and 150 (62.8%) patients showed the defect on 99mTc-MAA SPECT/CT. Among the accumulation group, 45 (50.5%) were classified as grade 1, 40 (44.9%) were grade 2, and 4 (4.5%) were grade 3. TLR gradually and significantly increased from grade 0 (0.009 ± 0.005) to grade 1 (0.021 ± 0.005, P < 0.05) and to grade 2-3 (0.033 ± 0.013, P < 0.05). The following factors were significant predictors for occult nodal metastasis in univariate analysis: central location, histology different from adenocarcinoma, tumor size greater than 3 cm representing clinical T2 or higher, and the absence of 99mTc-MAA accumulation within the tumor. Defect in the lung perfusion SPECT/CT remained significant at the multivariate analysis (Odd ratio 3.25, 95%CI [1.24 to 8.48], p = 0.016). With a median follow-up of 31.5 months, the RFS was significantly shorter in the defect group (p = 0.008). Univariate analysis revealed that cell type of non-adenocarcinoma, clinical stage II-III, pathologic stage II-III, age greater than 65 years, and the 99mTc-MAA defect within tumor as significant predictors for shorter RFS. However, only the pathologic stage remained statistically significant, in multivariate analysis. CONCLUSION: The absence of 99mTc-MAA accumulation within the tumor in preoperative lung perfusion SPECT/CT represents an independent risk factor for occult nodal metastasis and is relevant as a poor prognostic factor in clinically N0 NSCLC patients. 99mTc-MAA tumor distribution may serve as a new imaging biomarker reflecting tumor vasculatures and perfusion which can be associated with tumor biology and prognosis.

Mannosylated-serum albumin nanoparticle imaging to monitor tumor-associated macrophages under anti-PD1 treatment.

BACKGROUND: Immune checkpoint inhibitors such as anti-programmed cell death protein 1 (PD1) block tumor growth by reinvigorating the immune system; however, determining their efficacy only by the changes in tumor size may prove inaccurate. As the immune cells including macrophages in the tumor microenvironment (TME) are associated with the response to anti-PD1 therapy, tumor-associated macrophages (TAMs) imaging using nanoparticles can noninvasively provide the immune enrichment status of TME. Herein, the mannosylated-serum albumin (MSA) nanoparticle was labeled with radioactive isotope 68Ga to target the mannose receptors on macrophages for noninvasive monitoring of the TME according to anti-PD1 therapy. RESULTS: B16F10-Luc and MC38-Luc tumor-bearing mice were treated with anti-PD1, and the response to anti-PD1 was determined by the tumor volume. According to the flow cytometry, the responders to anti-PD1 showed an increased proportion of TAMs, as well as lymphocytes, and the most enriched immune cell population in the TME was also TAMs. For noninvasive imaging of TAMs as a surrogate of immune cell augmentation in the TME via anti-PD1, we acquired [68Ga] Ga-MSA positron emission tomography. According to the imaging study, an increased number of TAMs in responders at the early phase of anti-PD1 treatment was observed in both B16F10-Luc and MC38-Luc tumor-bearing mice models. CONCLUSION: As representative immune cells in the TME, non-invasive imaging of TAMs using MSA nanoparticles can reflect the immune cell enrichment status in the TME closely associated with the response to anti-PD1. As non-invasive imaging using MSA nanoparticles, this approach shows a potential to monitor and evaluate anti-tumor response to immune checkpoint inhibitors.

Discovery of molecular features underlying the morphological landscape by integrating spatial transcriptomic data with deep features of tissue images.

Profiling molecular features associated with the morphological landscape of tissue is crucial for investigating the structural and spatial patterns that underlie the biological function of tissues. In this study, we present a new method, spatial gene expression patterns by deep learning of tissue images (SPADE), to identify important genes associated with morphological contexts by combining spatial transcriptomic data with coregistered images. SPADE incorporates deep learning-derived image patterns with spatially resolved gene expression data to extract morphological context markers. Morphological features that correspond to spatial maps of the transcriptome were extracted by image patches surrounding each spot and were subsequently represented by image latent features. The molecular profiles correlated with the image latent features were identified. The extracted genes could be further analyzed to discover functional terms and exploited to extract clusters maintaining morphological contexts. We apply our approach to spatial transcriptomic data from different tissues, platforms and types of images to demonstrate an unbiased method that is capable of obtaining image-integrated gene expression trends.

Comparison of deep learning-based emission-only attenuation correction methods for positron emission tomography.

PURPOSE: This study aims to compare two approaches using only emission PET data and a convolution neural network (CNN) to correct the attenuation (µ) of the annihilation photons in PET. METHODS: One of the approaches uses a CNN to generate µ-maps from the non-attenuation-corrected (NAC) PET images (µ-CNNNAC). In the other method, CNN is used to improve the accuracy of µ-maps generated using maximum likelihood estimation of activity and attenuation (MLAA) reconstruction (µ-CNNMLAA). We investigated the improvement in the CNN performance by combining the two methods (µ-CNNMLAA+NAC) and the suitability of µ-CNNNAC for providing the scatter distribution required for MLAA reconstruction. Image data from 18F-FDG (n = 100) or 68 Ga-DOTATOC (n = 50) PET/CT scans were used for neural network training and testing. RESULTS: The error of the attenuation correction factors estimated using µ-CT and µ-CNNNAC was over 7%, but that of scatter estimates was only 2.5%, indicating the validity of the scatter estimation from µ-CNNNAC. However, CNNNAC provided less accurate bone structures in the µ-maps, while the best results in recovering the fine bone structures were obtained by applying CNNMLAA+NAC. Additionally, the µ-values in the lungs were overestimated by CNNNAC. Activity images (λ) corrected for attenuation using µ-CNNMLAA and µ-CNNMLAA+NAC were superior to those corrected using µ-CNNNAC, in terms of their similarity to λ-CT. However, the improvement in the similarity with λ-CT by combining the CNNNAC and CNNMLAA approaches was insignificant (percent error for lung cancer lesions, λ-CNNNAC = 5.45% ± 7.88%; λ-CNNMLAA = 1.21% ± 5.74%; λ-CNNMLAA+NAC = 1.91% ± 4.78%; percent error for bone cancer lesions, λ-CNNNAC = 1.37% ± 5.16%; λ-CNNMLAA = 0.23% ± 3.81%; λ-CNNMLAA+NAC = 0.05% ± 3.49%). CONCLUSION: The use of CNNNAC was feasible for scatter estimation to address the chicken-egg dilemma in MLAA reconstruction, but CNNMLAA outperformed CNNNAC.

Prediction of post-stroke cognitive impairment using brain FDG PET: deep learning-based approach.

PURPOSE: Post-stroke cognitive impairment can affect up to one third of stroke survivors. Since cognitive function greatly contributes to patients' quality of life, an objective quantitative biomarker for early prediction of dementia after stroke is required. We developed a deep-learning (DL)-based signature using positron emission tomography (PET) to objectively evaluate cognitive decline in patients with stroke. METHODS: We built a DL model that differentiated Alzheimer's disease (AD) from normal controls (NC) using brain fluorodeoxyglucose (FDG) PET from the Alzheimer's Disease Neuroimaging Initiative database. The model was directly transferred to a prospectively enrolled cohort of patients with stroke to differentiate patients with dementia from those without dementia. The accuracy of the model was evaluated by the area under the curve values of receiver operating characteristic curves (AUC-ROC). We visualized the distribution of DL-based features and brain regions that the model weighted for classification. Correlations between cognitive signature from the DL model and clinical variables were evaluated, and survival analysis for post-stroke dementia was performed in patients with stroke. RESULTS: The classification of AD vs. NC subjects was performed with AUC-ROC of 0.94 (95% confidence interval [CI], 0.89-0.98). The transferred model discriminated stroke patients with dementia (AUC-ROC = 0.75). The score of cognitive decline signature using FDG PET was positively correlated with age, neutrophil-lymphocyte ratio and platelet-lymphocyte ratio and negatively correlated with body mass index in patients with stroke. We found that the cognitive decline score was an independent risk factor for dementia following stroke (hazard ratio, 10.90; 95% CI, 3.59-33.09; P < 0.0001) after adjustment for other key variables. CONCLUSION: The DL-based cognitive signature using FDG PET was successfully transferred to an independent stroke cohort. It is suggested that DL-based cognitive evaluation using FDG PET could be utilized as an objective biomarker for cognitive dysfunction in patients with cerebrovascular diseases.

Image-level trajectory inference of tau pathology using variational autoencoder for Flortaucipir PET.

PURPOSE: Alzheimer's disease (AD) studies revealed that abnormal deposition of tau spreads in a specific spatial pattern, namely Braak stage. However, Braak staging is based on post mortem brains, each of which represents the cross section of the tau trajectory in disease progression, and numerous studies were reported that do not conform to that model. This study thus aimed to identify the tau trajectory and quantify the tau progression in a data-driven approach with the continuous latent space learned by variational autoencoder (VAE). METHODS: A total of 1080 [18F]Flortaucipir brain positron emission tomography (PET) images were collected from the Alzheimer's Disease Neuroimaging Initiative (ADNI) database. VAE was built to compress the hidden features from tau images in latent space. Hierarchical agglomerative clustering and minimum spanning tree (MST) were applied to organize the features and calibrate them to the tau progression, thus deriving pseudo-time. The image-level tau trajectory was inferred by continuously sampling across the calibrated latent features. We assessed the pseudo-time with regard to tau standardized uptake value ratio (SUVr) in AD-vulnerable regions, amyloid deposit, glucose metabolism, cognitive scores, and clinical diagnosis. RESULTS: We identified four clusters that plausibly capture certain stages of AD and organized the clusters in the latent space. The inferred tau trajectory agreed with the Braak staging. According to the derived pseudo-time, tau first deposits in the parahippocampal and amygdala, and then spreads to the fusiform, inferior temporal lobe, and posterior cingulate. Prior to the regional tau deposition, amyloid accumulates first. CONCLUSION: The spatiotemporal trajectory of tau progression inferred in this study was consistent with Braak staging. The profile of other biomarkers in disease progression agreed well with previous findings. We addressed that this approach additionally has the potential to quantify tau progression as a continuous variable by taking a whole-brain tau image into account.

Glucose metabolic profiles evaluated by PET associated with molecular characteristic landscape of gastric cancer.

BACKGROUND: Although FDG-PET is widely used in cancer, its role in gastric cancer (GC) is still controversial due to variable [18F]fluorodeoxyglucose ([18F]FDG) uptake. Here, we sought to develop a genetic signature to predict high FDG-avid GC to plan individualized PET and investigate the molecular landscape of GC and its association with glucose metabolic profiles noninvasively evaluated by [18F]FDG-PET. METHODS: Based on a genetic signature, PETscore, representing [18F]FDG avidity, was developed by imaging data acquired from thirty patient-derived xenografts (PDX). The PETscore was validated by [18F]FDG-PET data and gene expression data of human GC. The PETscore was associated with genomic and transcriptomic profiles of GC using The Cancer Genome Atlas. RESULTS: Five genes, PLS1, PYY, HBQ1, SLC6A5, and NAT16, were identified for the predictive model for [18F]FDG uptake of GC. The PETscore was validated in independent PET data of human GC with qRT-PCR and RNA-sequencing. By applying PETscore on TCGA, a significant association between glucose uptake and tumor mutational burden as well as genomic alterations were identified. CONCLUSION: Our findings suggest that molecular characteristics are underlying the diverse metabolic profiles of GC. Diverse glucose metabolic profiles may apply to precise diagnostic and therapeutic approaches for GC.

Glucose-Thymidine Ratio as a Metabolism Index Using 18F-FDG and 18F-FLT PET Uptake as a Potential Imaging Biomarker for Evaluating Immune Checkpoint Inhibitor Therapy.

Immune checkpoint inhibitors (ICIs) are widely used in cancer immunotherapy, requiring effective methods for response monitoring. This study evaluated changes in 18F-2-fluoro-2-deoxy-D-glucose (FDG) and 18F-fluorothymidine (FLT) uptake by tumors following ICI treatment as potential imaging biomarkers in mice. Tumor uptakes of 18F-FDG and 18F-FLT were measured and compared between the ICI treatment and control groups. A combined imaging index of glucose-thymidine uptake ratio (GTR) was defined and compared between groups. In the ICI treatment group, tumor growth was effectively inhibited, and higher proportions of immune cells were observed. In the early phase, 18F-FDG uptake was higher in the treatment group, whereas 18F-FLT uptake was not different. There was no difference in 18F-FDG uptake between the two groups in the late phase. However, 18F-FLT uptake of the control group was markedly increased compared with the ICI treatment group. GTR was consistently higher in the ICI treatment group in the early and late phases. After ICI treatment, changes in tumor cell proliferation were observed with 18F-FLT, whereas 18F-FDG showed altered metabolism in both tumor and immune cells. A combination of 18F-FLT and 18F-FDG PET, such as GTR, is expected to serve as a potentially effective imaging biomarker for monitoring ICI treatment.

Translating amyloid PET of different radiotracers by a deep generative model for interchangeability.

It is challenging to compare amyloid PET images obtained with different radiotracers. Here, we introduce a new approach to improve the interchangeability of amyloid PET acquired with different radiotracers through image-level translation. Deep generative networks were developed using unpaired PET datasets, consisting of 203 [11C]PIB and 850 [18F]florbetapir brain PET images. Using 15 paired PET datasets, the standardized uptake value ratio (SUVR) values obtained from pseudo-PIB or pseudo-florbetapir PET images translated using the generative networks was compared to those obtained from the original images. The generated amyloid PET images showed similar distribution patterns with original amyloid PET of different radiotracers. The SUVR obtained from the original [18F]florbetapir PET was lower than those obtained from the original [11C]PIB PET. The translated amyloid PET images reduced the difference in SUVR. The SUVR obtained from the pseudo-PIB PET images generated from [18F]florbetapir PET showed a good agreement with those of the original PIB PET (ICC = 0.87 for global SUVR). The SUVR obtained from the pseudo-florbetapir PET also showed a good agreement with those of the original [18F]florbetapir PET (ICC = 0.85 for global SUVR). The ICC values between the original and generated PET images were higher than those between original [11C]PIB and [18F]florbetapir images (ICC = 0.65 for global SUVR). Our approach provides the image-level translation of amyloid PET images obtained using different radiotracers. It may facilitate the clinical studies designed with variable amyloid PET images due to long-term clinical follow-up as well as multicenter trials by enabling the translation of different types of amyloid PET.

Hippocampal glucose uptake as a surrogate of metabolic change of microglia in Alzheimer's disease.

BACKGROUND: Dynamically altered microglia play an important role in the progression of Alzheimer's disease (AD). Here, we found a close association of the metabolic reconfiguration of microglia with increased hippocampal glucose uptake on [18F]fluorodeoxyglucose (FDG) PET. METHODS: We used an AD animal model, 5xFAD, to analyze hippocampal glucose metabolism using both animal FDG PET and ex vivo FDG uptake test. Cells of the hippocampus were isolated to perform single-cell RNA-sequencing (scRNA-seq). The molecular features of cells associated with glucose metabolism were analyzed at a single-cell level. In order to apply our findings to human brain imaging study, brain FDG PET data obtained from the Alzheimer's Disease Neuroimaging Initiative were analyzed. FDG uptake in the hippocampus was compared according to the diagnosis, AD, mild cognitive impairment, and controls. The correlation analysis between hippocampal FDG uptake and soluble TREM2 in cerebrospinal fluid was performed. RESULTS: In the animal study, 8- and 12-month-old 5xFAD mice showed higher FDG uptake in the hippocampus than wild-type mice. Cellular FDG uptake tests showed that FDG activity in hippocampal microglia was increased in the AD model, while FDG activity in non-microglial cells of the hippocampus was not different between the AD model and wild-type. scRNA-seq data showed that changes in glucose metabolism signatures including glucose transporters, glycolysis and oxidative phosphorylation, mainly occurred in microglia. A subset of microglia with higher glucose transporters with defective glycolysis and oxidative phosphorylation was increased according to disease progression. In the human imaging study, we found a positive association between soluble TREM2 and hippocampal FDG uptake. FDG uptake in the hippocampus at the baseline scan predicted mild cognitive impairment conversion to AD. CONCLUSIONS: We identified the reconfiguration of microglial glucose metabolism in the hippocampus of AD, which could be evaluated by FDG PET as a feasible surrogate imaging biomarker for microglia-mediated inflammation.

Visual interpretation of [18F]Florbetaben PET supported by deep learning-based estimation of amyloid burden.

PURPOSE: Amyloid PET which has been widely used for noninvasive assessment of cortical amyloid burden is visually interpreted in the clinical setting. As a fast and easy-to-use visual interpretation support system, we analyze whether the deep learning-based end-to-end estimation of amyloid burden improves inter-reader agreement as well as the confidence of the visual reading. METHODS: A total of 121 clinical routines [18F]Florbetaben PET images were collected for the randomized blind-reader study. The amyloid PET images were visually interpreted by three experts independently blind to other information. The readers qualitatively interpreted images without quantification at the first reading session. After more than 2-week interval, the readers additionally interpreted images with the quantification results provided by the deep learning system. The qualitative assessment was based on a 3-point BAPL score (1: no amyloid load, 2: minor amyloid load, and 3: significant amyloid load). The confidence score for each session was evaluated by a 3-point score (0: ambiguous, 1: probably, and 2: definite to decide). RESULTS: Inter-reader agreements for the visual reading based on a 3-point scale (BAPL score) calculated by Fleiss kappa coefficients were 0.46 and 0.76 for the visual reading without and with the deep learning system, respectively. For the two reading sessions, the confidence score of visual reading was improved at the visual reading session with the output (1.27 ± 0.078 for visual reading-only session vs. 1.66 ± 0.63 for a visual reading session with the deep learning system). CONCLUSION: Our results highlight the impact of deep learning-based one-step amyloid burden estimation system on inter-reader agreement and confidence of reading when applied to clinical routine amyloid PET reading.

Unsupervised clustering of dopamine transporter PET imaging discovers heterogeneity of parkinsonism.

Parkinsonism has heterogeneous nature, showing distinctive patterns of disease progression and prognosis. We aimed to find clusters of parkinsonism based on 18 F-fluoropropyl-carbomethoxyiodophenylnortropane (FP-CIT) PET as a data-driven approach to evaluate heterogenous dopaminergic neurodegeneration patterns. Two different cohorts of patients who received FP-CIT PET were collected. A labeled cohort (n = 94) included patients with parkinsonism who underwent a clinical follow-up of at least 3 years (mean 59.0 ± 14.6 months). An unlabeled cohort (n = 813) included all FP-CIT PET data of a single-center. All PET data were clustered by a dimension reduction method followed by hierarchical clustering. Four distinct clusters were defined according to the imaging patterns. When the diagnosis of the labeled cohort of 94 patients was compared with the corresponding cluster, parkinsonism patients were mostly included in two clusters, cluster "0" and "2." Specifically, patients with progressive supranuclear palsy were significantly more included in cluster 0. The two distinct clusters showed significantly different clinical features. Furthermore, even in PD patients, two clusters showed a trend of different clinical features. We found distinctive clusters of parkinsonism based on FP-CIT PET-derived heterogeneous neurodegeneration patterns, which were associated with different clinical features. Our results support a biological underpinning for the heterogeneity of neurodegeneration in parkinsonism.

Deep learning-based interpretation of basal/acetazolamide brain perfusion SPECT leveraging unstructured reading reports.

PURPOSE: Basal/acetazolamide brain perfusion single-photon emission computed tomography (SPECT) has been used to evaluate functional hemodynamics in patients with carotid artery stenosis. We aimed to develop a deep learning model as a support system for interpreting brain perfusion SPECT leveraging unstructured text reports. METHODS: In total, 7345 basal/acetazolamide brain perfusion SPECT images and their text reports were retrospectively collected. A long short-term memory (LSTM) network was trained using 500 randomly selected text reports to predict manually labeled structured information, including abnormalities of basal perfusion and vascular reserve for each vascular territory. Using this trained LSTM model, we extracted structured information from the remaining 6845 text reports to develop a deep learning model for interpreting SPECT images. The model was based on a 3D convolutional neural network (CNN), and the performance was tested on the other 500 cases by measuring the area under the receiver-operating characteristic curve (AUC). We then applied the model to patients who underwent revascularization (n = 33) to compare the estimated output of the CNN model for pre- and post-revascularization SPECT and clinical outcomes. RESULTS: The AUC of the LSTM model for extracting structured labels was 1.00 for basal perfusion and 0.99 for vascular reserve for all 9 brain regions. The AUC of the CNN model designed to identify abnormal perfusion was 0.83 for basal perfusion and 0.89 for vascular reserve. The output of the CNN model was significantly improved according to the revascularization in the target vascular territory, and its changes in brain territories were concordant with clinical outcomes. CONCLUSION: We developed a deep learning model to support the interpretation of brain perfusion SPECT by converting unstructured text reports into structured labels. This model can be used as a support system not only to identify perfusion abnormalities but also to provide quantitative scores of abnormalities, particularly for patients who require revascularization.

Cognitive signature of brain FDG PET based on deep learning: domain transfer from Alzheimer's disease to Parkinson's disease.

PURPOSE: Although functional brain imaging has been used for the early and objective assessment of cognitive dysfunction, there is a lack of generalized image-based biomarker which can evaluate individual's cognitive dysfunction in various disorders. To this end, we developed a deep learning-based cognitive signature of FDG brain PET adaptable for Parkinson's disease (PD) as well as Alzheimer's disease (AD). METHODS: A deep learning model for discriminating AD from normal controls (NCs) was built by a training set consisting of 636 FDG PET obtained from Alzheimer's Disease Neuroimaging Initiative database. The model was directly transferred to images of mild cognitive impairment (MCI) patients (n = 666) for identifying who would rapidly convert to AD and another independent cohort consisting of 62 PD patients to differentiate PD patients with dementia. The model accuracy was measured by area under curve (AUC) of receiver operating characteristic (ROC) analysis. The relationship between all images was visualized by two-dimensional projection of the deep learning-based features. The model was also designed to predict cognitive score of the subjects and validated in PD patients. Cognitive dysfunction-related regions were visualized by feature maps of the deep CNN model. RESULTS: AUC of ROC for differentiating AD from NC was 0.94 (95% CI 0.89-0.98). The transfer of the model could differentiate MCI patients who would convert to AD (AUC = 0.82) and PD with dementia (AUC = 0.81). The two-dimensional projection mapping visualized the degree of cognitive dysfunction compared with normal brains regardless of different disease cohorts. Predicted cognitive score, an output of the model, was highly correlated with the mini-mental status exam scores. Individual cognitive dysfunction-related regions included cingulate and high frontoparietal cortices, while they showed individual variability. CONCLUSION: The deep learning-based cognitive function evaluation model could be successfully transferred to multiple disease domains. We suggest that this approach might be extended to an objective cognitive signature that provides quantitative biomarker for cognitive dysfunction across various neurodegenerative disorders.

Comprehensive gene expression analysis for exploring the association between glucose metabolism and differentiation of thyroid cancer.

BACKGROUND: The principle of loss of iodine uptake and increased glucose metabolism according to dedifferentiation of thyroid cancer is clinically assessed by imaging. Though these biological properties are widely applied to appropriate iodine therapy, the understanding of the genomic background of this principle is still lacking. We investigated the association between glucose metabolism and differentiation in advanced thyroid cancer as well as papillary thyroid cancer (PTC). METHODS: We used RNA sequencing of 505 patients with PTC obtained from the Cancer Genome Archives and microarray data of poorly-differentiated and anaplastic thyroid cancer (PDTC/ATC). The signatures of GLUT and glycolysis were estimated to assess glucose metabolic profiles. The glucose metabolic profiles were associated with tumor differentiation score (TDS) and BRAFV600E mutation status. In addition, survival analysis of glucose metabolic profiles was performed for predicting recurrence-free survival. RESULTS: In PTC, the glycolysis signature was positively correlated with TDS, while the GLUT signature was inversely correlated with TDS. These correlations were significantly stronger in the BRAFV600E negative group than the positive group. Meanwhile, both GLUT and glycolysis signatures were negatively correlated with TDS in advanced thyroid cancer. The high glycolysis signature was significantly associated with poor prognosis in PTC in spite of high TDS. The glucose metabolic profiles are intricately associated with tumor differentiation in PTC and PDTC/ATC. CONCLUSIONS: As glycolysis was an independent prognostic marker, we suggest that the glucose metabolism features of thyroid cancer could be another biological progression marker different from differentiation and provide clinical implications for risk stratification. TRIAL REGISTRATION: Not applicable.