Greg Zaharchuk

Background: Gadolinium‐based contrast agents are used in brain MRI to improve the visualization of disorders and improve the delineation of lesions. Higher doses of GBCAs can improve lesion sensitivity but may have safety implications, particularly in light of recent findings on gadolinium retention and deposition. Purpose: To evaluate the clinical performance of an FDA‐cleared deep‐learning (DL)‐based contrast boosting algorithm in routine clinical brain MRI exams. Study Type: Retrospective. Population: One hundred ten patients (47 ± 22 years; 52 Females, 47 Males, 11 N/A) with clinical contrast‐enhanced brain MR studies. Field Strength and Sequences: T1 weighted pre‐contrast and post‐contrast brain MRI sequences at 0.3 T, 1.5 T, and 3 T. Assessment: A multi‐center database of contrast‐enhanced brain MR images was used to evaluate a DL‐based contrast boosting algorithm. Pre‐contrast and standard post‐contrast (SC) images were processed with the algorithm to obtain contrast boosted (CB) images. CB images were compared to SC images in terms of contrast‐to‐noise ratio (CNR), lesion‐to‐brain ratio (LBR), and contrast enhancement percentage (CEP). Three board‐certified radiologists with 9, 15, and 18 years of experience reviewed CB and SC images side‐by‐side for qualitative evaluation and rated them on a 4‐point Likert scale for lesion contrast enhancement, border delineation, internal morphology, overall image quality, presence of artifacts, and changes in vessel conspicuity. The presence, cause, and severity of any false lesions was recorded. Statistical Tests: Wilcoxon signed rank test. A p value < 0.05 was considered significant. Results: CB images had significantly superior quantitative performance than SC images in terms of CNR (729.17% ± 1576.92%), LBR (87.91% ± 72.12%), and CEP (165.81% ± 42%). In the qualitative assessment, CB images showed significantly better lesion visualization (3.73 vs. 3.16) and had significantly better image quality (3.55 vs. 3.07). Data Conclusion: In this multi‐center, multi‐reader study, deep learning–based contrast boosting demonstrates robust improvements in lesion visualization and image quality without increasing contrast dosage. Evidence Level: 4. Technical Efficacy: Stage 3.

Background: Deep learning (DL)‐based image enhancement is widely used to improve suboptimal medical imaging. Whether it also benefits diagnostic‐quality MRI in downstream task performance and data‐efficiency remains unclear. Purpose: To investigate the impact of DL‐based enhancement applied to diagnostic quality structural MRI for Alzheimer's disease (AD) classification. Study Type: Retrospective. Population: A total of 2293 brain MRI scans from the Alzheimer's Disease Neuroimaging Initiative (ADNI) were split into training ( n = 1605), validation ( n = 229), and internal test ( n = 459) sets. Two hundred and seventy scans from the National Alzheimer's Coordinating Center (NACC) served as an external test set. Field Strength/Sequence: 1.5 T; 3D T1‐weighted gradient‐echo. Assessment: Each scan was enhanced by SubtleHD (SHD), an FDA‐cleared DL‐based MR enhancement tool. ResNet34 and DenseNet121 were trained on standard‐of‐care (SOC) and SHD‐enhanced images to classify subjects as cognitively normal, mild cognitive impairment, or AD and evaluated by accuracy and macro–area under the receiver operating characteristic curve (macro‐AUC). Data efficiency was assessed by retraining on stratified training subsets (50%–100%). Statistical Tests: McNemar test for accuracy and DeLong test for macro‐AUC in three‐class one‐versus‐rest setting ( p < 0.05). Results: SHD enhancement increased ResNet34 accuracy from 85.2% to 88.7% and macro‐AUC from 0.951 to 0.968 (both significant), and DenseNet121 accuracy from 90.2% to 92.2% ( p = 0.18) and macro‐AUC from 0.978 to 0.982 ( p = 0.29). Models trained on 70% of SHD‐enhanced dataset matched those trained on the full SOC dataset (accuracy: 85.9%, macro‐AUC: 0.942), indicating improved data efficiency with enhancement. In NACC, the SOC‐trained model achieved accuracy of 49.2% and macro‐AUC of 0.679 versus 63.0% and 0.819 for the SHD‐trained model (both significant); the SHD‐trained model retained an advantage on unenhanced NACC images (macro‐AUC: 0.772). Data Conclusion: DL‐based enhancement of diagnostic‐quality MRI improves downstream Alzheimer's disease classification performance and reduces the amount of training data required. This suggests that conventional definitions of image quality may underestimate the information content available for machine learning. Evidence Level: 3. Technical Efficacy Stage: 2.

Fluorodeoxyglucose (FDG) PET to evaluate patients with epilepsy is one of the most common applications for simultaneous PET/MRI, given the need to image both brain structure and metabolism but is suboptimal due to the radiation dose in this young population. Little work has been done synthesizing diagnostic quality PET images from MRI data or MRI data with ultralow-dose PET using advanced generative AI methods, such as diffusion models, with attention to clinical evaluations tailored for the epilepsy population. We compared the performance of diffusion- and non-diffusion-based deep learning models for the MRI-to-PET image translation task for epilepsy imaging using simultaneous PET/MRI in 52 subjects (40 train/2 validate/10 hold-out test). We tested three different models: 2 score-based generative diffusion models (SGM-Karras Diffusion [SGM-KD] and SGM-variance preserving [SGM-VP]) and a Transformer-U-net. We report results on standard image processing metrics as well as clinically relevant metrics, including congruency measures (Congruence Index and Congruency Mean Absolute Error) that assess hemispheric metabolic asymmetry, which is a key part of the clinical analysis of these images. We compared the model performance using different inputs such as T1-weighted (T1w), T2 FLAIR (T2F), and 1% ultralow-dose PET images to evaluate the effect and necessity of each imaging contrast. The SGM-KD produced the best qualitative and quantitative results when synthesizing PET purely from T1w and T2 FLAIR images with the least mean absolute error in whole-brain specific uptake value ratio (SUVR) and highest intraclass correlation coefficient. When 1% low-dose PET images are included in the inputs, all models improve significantly and are interchangeable for quantitative performance and visual quality. SGMs hold great potential for pure MRI-to-PET translation, while all 3 model types can synthesize full-dose FDG-PET accurately using MRI and ultralow-dose PET. This suggests that deep learning diffusion models could reduce or eliminate radiation dose for patients being evaluated for epilepsy.

Background: Deep learning (DL) based image enhancement is widely used to improve suboptimal medical imaging. However, whether such techniques can extract additional diagnostically relevant information from already high quality, clinically acceptable MRI, and thereby improve downstream artificial intelligence (AI) model performance and data-efficiency, remains unclear. We investigated the impact of DL-based enhancement on AI-based Alzheimer’s disease (AD) classification from structural MRI. Methods: In this retrospective study, we included 2,293 MRI scans from the public Alzheimer’s Disease Neuroimaging Initiative dataset (training/validation/test 7:1:2), upon which multiple AI classification models were trained, first on the original images (standard-of-care [SOC]), then on the same images enhanced by SubtleHD (SHD), an FDA-cleared enhancement tool. Convolutional neural networks (ResNet34 and DenseNet121) were used to classify subjects as cognitively normal (CN), mild cognitive impairment (MCI), or AD. Model performance was assessed using accuracy and macro–area under the receiver operating characteristic curve (macro-AUC). Data efficiency was assessed by training on progressively smaller subsets of the full dataset. Findings: Models trained and evaluated on enhanced images consistently outperformed those using the original images. For ResNet34, enhancement improved classification accuracy from 85·2% to 88·7% and macro-AUC from 0·951 to 0·968 (both p<0·01). Similar gains were observed with DenseNet121. Notably, models trained on 75% of SHD-enhanced images achieved comparable performance to models trained on the full SOC dataset (accuracy 86·5%, macro-AUC 0·954), indicating improved data efficiency.Interpretations: DL-based enhancement of MRI image quality, even when images are already considered to be of diagnostic quality, improves downstream AI classification performance and reduces the amount of training data required. This has profound implications for radiologic imaging, suggesting that conventional definitions of image quality may underestimate the information content available for machine learning. DL-based enhancement strategies could also enable more efficient development and broader deployment of AI models in neuroimaging, particularly in data-limited settings.

Deep learning enables the extraction of ischemic lesion size and hypodensity imaging markers from noncontrast CT (DLNCCT) in patients with acute ischemic stroke, but it remains unclear whether those markers can predict post-transfer core volume. We performed a post-hoc analysis of prospectively enrolled patients transferred from a primary to a comprehensive center (PSC/CSC) for endovascular treatment (EVT). Using a validated deep-learning NCCT segmentation method, we quantified total lesion volume (per 10mL), modified net water uptake (mNWU %), and severely hypodense volume (≤26 HU per 10mL) and compared these markers with core-lab-rated ASPECTS (per point decrease) and CTP-based evaluation for their association with (adjusted regression coefficient [95%CI]) and predictive performance in addition to baseline variables (R2±SE) for post-transfer CSC-admission DWI core volume. We included 420 patients (239[57%] males) with a median age of 72 years (IQR:61;80). We observed 11.2mL (95%CI:8.3;14.1] larger post-transfer core volumes per point decrease in ASPECTS, 10.0mL (95%CI:6.8;13.3) and 20.0mL (95%CI:12.7;27.2) larger post-transfer core volumes per 10 mL increase in total and severely hypodense DLNCCT volume, respectively. mNWU was not associated with post-transfer core volume (p=0.63). In addition to clinical baseline and CTA variables, post-transfer core volume prediction with ASPECTS (R2:0.49±0.02) and DLNCCT (R2:0.50±0.02) did not differ significantly (p=0.58). Compared with using CTP imaging markers (R2:0.56±0.02), adding ASPECTS (R2:0.63±0.02, p<0.01) and DLNCCT (R2:0.65±0.01, p<0.01) improved performance for post-transfer core volume prediction. Total and severely hypodense DLNCCT volumes are independent predictors for post-transfer core volume. These DLNCCT markers improved CTP-based post-transfer core volume prediction.

18F-Florbetaben (FBB) uptake in the supratentorial cortex is indicative of amyloid positivity. Due to PET’s low spatial resolution, image noise, and spill-over of signal from adjacent white-matter into gray-matter, trained readers may provide inconsistent reads and quantitative calculations like Centiloids (CLs) are also affected. A set of 264 18F-Florbetaben (amyloid) PET/MRI exams were reconstructed using conventional ordered subset expectation maximization (OSEM) method and MRguided block sequential regularized expectation maximization (MRgBSREM) method. Three trained readers evaluated the images from these 264 patients, which were reconstructed using the OSEM method. Fifty-three exams were rated inconsistently and were mixed with another 53 exams which were rated consistently. These 106 subjects were then rated by our readers using the MRgBSREM PET reconstruction method. CLs were measured using both reconstruction methods. There is significant correlation between CL measured by OSEM and MRgBSREM methods with R2=0.99. The number of inconsistent exams dropped by 64% using MRgBSREM method as compared with OSEM method. Using Fleiss-Kappa statistical test, the agreement between readers was raised from “Fair” to “Significant” in the 106-subjects subset. PET reconstruction with MR priors can significantly improve the consistency of ratings among trained readers. Given the prevalence of inconsistent ratings in amyloid PET, methods that enhance the ability to distinguish intermediate amyloid levels could be valuable for the widespread adoption of this modality.

Artificial intelligence (AI) models have shown promise in neuroradiology, yet their real-world generalizability remains uncertain, partly due to variability in imaging acquisition and protocols. We aimed to evaluate the impact of data source, scanner manufacturer, scan mode, slice thickness, and the AI models developed by participating teams on AI performance in this secondary analysis of the 2019 American Society of Functional Neuroradiology (ASFNR) AI Competition. We included 1,177 anonymized noncontrast head CT scans from five institutions. Four teams participated, developing models to detect acute ischemic stroke, intracranial hemorrhage, mass effect, and to assess age-appropriate normality. Generalized estimating equations (GEE) were used to evaluate the effects of the aforementioned variables on model performance, and collinearity diagnostics were applied to exclude redundant variables. Due to collinearity with scanner manufacturer, data source was excluded from the model. Across all tasks, the AI model employed significantly influenced performance. Scanner manufacturer was significantly associated with accuracy in detecting intracranial hemorrhage and acute ischemic stroke but not mass effect or age-based normality. Slice thickness significantly associated with detection of intracranial hemorrhage and mass effect, with thinner slices yielding higher accuracy, but showed no effect on ischemic stroke or normality assessments. Scan mode did not significantly influence performance for any task. This secondary analysis demonstrates that imaging acquisition and protocol variability may significantly affect AI model performance. Scanner manufacturer, slice thickness, and the developed AI model were significantly associated with model accuracy, whereas scan mode had no significant impact. Among these, the developed AI model consistently proved most influential, reflecting the importance of training data, model architecture, and preprocessing methods.

Background: Multi‐center imaging studies create large‐scale data that are useful for identifying pathological patterns and robust training of deep learning models. However, variation due to site and scanner differences can confound analyses, emphasizing the need for harmonization. Purpose: To evaluate scanner‐related differences in T1w and T2‐FLAIR images in the Alzheimer's Disease Neuroimaging Initiative (ADNI) dataset and assess the performance of publicly available image‐level harmonization tools. Study Type: Retrospective. Population: Scanner group analysis: 1143 ADNI3 subjects (233 GE, 173 Philips, 250 Siemens, with 487 Siemens subjects used as an independent reference group). Within‐subject comparison: paired multi‐vendor scan sessions from 8 subjects. Field Strength/Sequence: 3.0T, T1w, and T2‐FLAIR MRI sequences. Assessment: Gray/white matter contrast ratio (G/W ratio), white matter hyperintensity (WMH) volume, and image feature similarity metrics (Fréchet Inception Distance [FID], Learned Perceptual Image Patch Similarity [LPIPS]) were compared across scanner vendors before and after harmonization with statistical (ComBat) and deep learning (HACA3) algorithms. Statistical Tests: One‐way ANOVA and post hoc Games‐Howell tests were conducted to assess differences between scanner groups across image pipelines (baseline, post‐harmonization). Repeated‐measures ANOVA and post hoc paired t ‐tests with Bonferroni correction were used to evaluate similarity metric changes pre‐ and post‐harmonization for multi‐vendor subjects. We defined statistical significance as p < 0.05. Results: At baseline, significant image differences in G/W ratio and WMH volumes between vendors were identified. Both ComBat and HACA3 harmonization improved G/W ratio consistency for T1w and T2‐FLAIR imaging across vendors, particularly for GE T2‐FLAIRs. HACA3 led to the best similarity between scanner datasets: mean FID T1w/T2‐FLAIR: 10.45/14.62 (Baseline); 7.45/11.71 (ComBat); 5.60/8.91 (HACA3). Only HACA3 harmonization resulted in non‐significant differences between vendors for WMH volume. Data Conclusion: HACA3 deep learning harmonization outperformed a statistical method, ComBat, improving MR contrast consistency and feature similarity across vendors. However, difficulties in harmonizing T2‐FLAIRs highlight limitations in current multi‐contrast MR harmonization tools. Evidence Level: 3. Technical Efficacy: Stage 1.

BACKGROUND AND PURPOSE: Predicting long-term clinical outcomes based on early acute ischemic stroke (AIS) information would be useful for many reasons, including patient counseling and clinical trial execution. This study investigates how different regions in brain imaging, including noninfarcted areas, contribute to the accuracy of predicting 90-day stroke outcomes by using deep learning (DL). MATERIALS AND METHODS: We developed and validated DL models in 449 patients with AIS, by using MRI DWI scans from 1–7 days poststroke and 90-day mRS outcome data. These models were trained on various inputs: infarct volumes, full-brain images, infarct masks, intensity-preserved infarct masks, and images in which the infarct region is removed, which we call lesion-neutralized images. Performance was assessed by using accuracy of predicting the specific mRS score, accuracy within ±1 mRS category, mean absolute error (MAE), and the area under the curve (AUC) to predict unfavorable outcome (mRS > 2). RESULTS: The model trained by using only infarct volume size reported the highest (worst) MAE of 1.51 points (95% CI, 1.40–1.61; P < .001), while the model trained with full-brain images achieved the lowest MAE of 1.07 points (95% CI, 0.99–1.16). Models with intermediate amounts of imaging information each improved on the volume-only predictions but did not reach the performance of the full brain images; infarct masks, intensity-preserved infarct masks, and lesion-neutralized images demonstrated MAEs of 1.25 (95% CI, 1.15–1.34; P = .002), 1.21 (95% CI, 1.11–1.30; P = .008), and 1.35 (95% CI, 1.24–1.45; P < .001), respectively. Similar results were seen for other prediction tasks, including AUC to predict unfavorable outcomes, ranging from 0.68 (95% CI, 0.63–0.73) for infarct volume to 0.86 (95% CI, 0.82–0.89) for full brain inputs. CONCLUSIONS: While the best performance came from by using the full brain imaging volume, we demonstrate that the infarct location, its signal characteristics, and importantly, the noninfarcted regions all contribute to the predictions. The noninfarcted areas may be a proxy for overall brain health and resilience, containing important information about potential outcomes.

With the rapid dissemination of artificial intelligence acceleration tools for MRI, individual radiologists and health systems should critically and objectively evaluate the effectiveness of their use in clinical practice.