Faculty Bibliography

Background: Increasing evidence suggests the subtle changes of Blood-Brain Barrier (BBB) permeability in normal aging and in Alzheimer"™s disease using Dynamic Contrast-Enhanced MRI (DCE-MRI). However, measuring this subtle change poses great challenge for accurate measurement, resulting in inconsistent results among previous studies. Two major challenges are long scan times, as suggested by previous studies and selection of the arterial input function (AIF). In this study, we aim to estimate the capillary level input function (CIF) using a deep learning network to overcome these two challenges. Methods: Healthy volunteers (n= 8, ages: 21-76) were recruited for DCE-MRI scan for 28min. Golden-angle RAdial Sampling Parallel (GRASP) sequence was used to obtain the dynamic images at ∼5s/frame. Individual AIF was sampled from the superior sagittal sinus of the brain (Fig.1a). FSL was used to segment the gray and white matters (Fig.1b). Each voxel was fitted using the graphical Patlak model (Fig.2a) to assess the vascular permeability-surface area product (PS) for both 28-min data and 10-min truncated data. We used a 3x3 kernel sliding through the images (Fig.3) and feed each voxel"™s dynamic as the input to our vision-transformer. Training data were generated using individual AIFs with a mathematical model, consisting of two Gaussian and one exponential function, and used to simulate dynamic patches using the Extended Patlak model (Fig.2b). Result: When the 10-min data are used, the conventional approach with AIF results in overestimation of PS when the scan-time is reduced, while the network-predicted CIF allows more accurate estimation, with refence to the results using the 28-min data, as illustrated by an example in Figure 4. Figure 5 shows the regional permeability differences between young and old subjects, where the conventional approach with AIF does not show the difference, while the approach with CIF shows subtle increases in PS with aging. Conclusion: Our proposed CIF-based approach provides an appropriate input-function for DCE analysis, allowing assessment of subtle permeability changes in the BBB.

PURPOSE/OBJECTIVE:The aim of this study was to investigate the feasibility of using 2 flip angles (FAs) with an ultrashort echo time during dynamic contrast-enhanced (DCE)-magnetic resonance imaging (MRI) for estimation of plasma gadolinium (Gd) concentration without using a precontrast longitudinal relaxation time T1 (T10) measurement. METHODS:T1-weighted DCE-MRI experiments were carried out with C57BL/6J mice using the scan protocol with 2 FAs over 3 sequential segments during 1 scan. The data with 2 FAs were used to estimate T10 (T1T) during conversion of a time-intensity curve to the time-concentration curve. Three dosages of gadolinium-based contrast agent were used to achieve a wide range of variability in Gd concentrations when measured at 10 minutes postinjection: 0.05 mmol/kg (n = 6), 0.1 mmol/kg (n = 11), and 0.15 mmol/kg (n = 7). For comparison, the signal-to-concentration conversion was also conducted using the T10 measured from the precontrast scan (T1M) as well as a constant T10 (2.1 seconds) from the literature (T1C). The Gd concentrations ([Gd]) estimated using DCE-MRI data for the time of retro-orbital blood collection ([Gd]T1T, [Gd]T1M, and [Gd]T1C, respectively) were compared against the [Gd] of the blood samples measured by inductively coupled plasma mass spectrometry ([Gd]MS). In addition, contrast kinetic model analysis was conducted on mice with GL261 brain tumors (n = 5) using the 3 different methods for T10. RESULTS:T1T strongly correlated with T1M (r = 0.81). [Gd]T1M and [Gd]T1T were significantly different from [Gd]T1C. [Gd]T1M and [Gd]T1T were in good agreement with [Gd]MS with strong correlations (mean percentage error ± standard deviation) of r = 0.70 (16% ± 56%) and r = 0.85 (15% ± 44%), respectively. In contrast, [Gd]T1C had a weak correlation of r = 0.52 with larger errors of 33% ± 24%. The contrast kinetic model parameters of GL261 brain tumors using T1T were not significantly different from those using T1M. CONCLUSIONS:This study substantiates the feasibility of using the 2-FA approach during DCE-MRI scan to estimate [Gd] in the plasma without using an extra scan to perform precontrast T1 measurements.

PURPOSE/OBJECTIVE:To develop a deep learning approach to estimate the local capillary-level input function (CIF) for pharmacokinetic model analysis of DCE-MRI. METHODS:A deep convolutional network was trained with numerically simulated data to estimate the CIF. The trained network was tested using simulated lesion data and used to estimate voxel-wise CIF for pharmacokinetic model analysis of breast DCE-MRI data using an abbreviated protocol from women with malignant (n = 25) and benign (n = 28) lesions. The estimated parameters were used to build a logistic regression model to detect the malignancy. RESULT/RESULTS:The pharmacokinetic parameters estimated using the network-predicted CIF from our breast DCE data showed significant differences between the malignant and benign groups for all parameters. Testing the diagnostic performance with the estimated parameters, the conventional approach with arterial input function (AIF) showed an area under the curve (AUC) between 0.76 and 0.87, and the proposed approach with CIF demonstrated similar performance with an AUC between 0.79 and 0.81. CONCLUSION/CONCLUSIONS:This study shows the feasibility of estimating voxel-wise CIF using a deep neural network. The proposed approach could eliminate the need to measure AIF manually without compromising the diagnostic performance to detect the malignancy in the clinical setting.

Gadolinium (Gd) based contrast agents (GBCAs) are widely used in magnetic resonance imaging (MRI) and are paramount to cancer diagnostics and tumor pharmacokinetic analysis. Accurate quantification of gadolinium concentration is essential to monitoring the biodistribution, clearance, and pharmacodynamics of GBCAs. However, current methods of quantifying gadolinium in blood or plasma (biological media) are both low throughput and clinically unavailable. Here, we have demonstrated the use of a sensitized gadolinium chelate, Gd[DTPA-cs124], as an MRI contrast agent that can be used to measure the concentration of gadolinium via luminescence quantification in biological media following transmetalation with a terbium salt. Gd[DTPA-cs124] was synthesized by conjugating carbostyril-124 (cs124) to diethylenetriaminepentaacetic acid (DTPA) and chelating to gadolinium. We report increases in both stability and relaxivity compared to the clinically approved analog Gd[DTPA] (gadopentetic acid or Magnevist). In vivo MRI experiments were conducted using C57BL6 mice in order to further illustrate the performance of Gd[DTPA-cs124] as an MRI contrast agent in comparison to Magnevist. Our results indicate that similar chemical modification to existing clinically approved GBCA may likewise provide favorable property changes, with the ability to be used in a gadolinium quantification assay. Furthermore, our assay provides a straightforward and high-throughput method of measuring gadolinium in biological media using a standard laboratory plate reader.

PURPOSE/OBJECTIVE:To develop a simultaneous dual-slab three-dimensional gradient-echo spectroscopic imaging (GSI) technique with frequency drift compensation for rapid (<6 min) bilateral measurement of fatty acid composition (FAC) in mammary adipose tissue. METHODS:A bilateral GSI sequence was developed using a simultaneous dual-slab excitation followed by 128 monopolar echoes. A short train of navigator echoes without phase or partition encoding was included at the beginning of each pulse repetition time period to correct for frequency variation caused by respiration and heating of the cryostat. Voxel-wise spectral fitting was applied to measure the areas of the lipid spectral peaks to estimate the number of double-bond (ndb), number of methylene-interrupted double-bond (nmidb), and chain length (cl). The proposed method was tested in an oil phantom and 10 postmenopausal women to assess the influence of the frequency variation on FAC estimation. RESULTS:The frequency drift observed over 5:27 min during the phantom scan was about 10 Hz. Phase correction based on the navigator reduced the median error of ndb, nmidb, and cl from 9.7%, 17.6%, and 3.2% to 2.1%, 9.5%, and 2.8%, respectively. The in vivo data showed a mean ± standard deviation frequency drift of 17.4 ± 2.5 Hz, with ripples at 0.3 ± 0.1 Hz. Our reconstruction algorithm successfully separated signals from the left and right breasts with negligible residual aliasing. Phase correction reduced the interquartile range within each subject's adipose tissue of ndb, nmidb, and cl by 18.4 ± 10.6%, 18.5 ± 13.9%, and 18.4 ± 10.6%, respectively. CONCLUSION/CONCLUSIONS:This study shows the feasibility of obtaining bilateral spectroscopic imaging data in the breast and that incorporation of a frequency navigator improves the estimation of FAC.

PURPOSE/OBJECTIVE:) in tumors, both in animals and in humans. METHODS:) by adjusting the diffusion gradient strength. The tDKI data at each diffusion time t were used for a weighted linear least-squares fit method to estimate the diffusion-time-dependent diffusivity, D(t), and diffusional kurtosis, K(t). RESULTS:median and IQR of the two breast cancers were 70 (94) and 106 (92) ms. CONCLUSION/CONCLUSIONS:The results of this proof-of-concept study substantiate the feasibility of using tDKI to measure cellular-interstitial water exchange time without using an exogenous contrast agent.

Medical images differ from natural images in significantly higher resolutions and smaller regions of interest. Because of these differences, neural network architectures that work well for natural images might not be applicable to medical image analysis. In this work, we propose a novel neural network model to address these unique properties of medical images. This model first uses a low-capacity, yet memory-efficient, network on the whole image to identify the most informative regions. It then applies another higher-capacity network to collect details from chosen regions. Finally, it employs a fusion module that aggregates global and local information to make a prediction. While existing methods often require lesion segmentation during training, our model is trained with only image-level labels and can generate pixel-level saliency maps indicating possible malignant findings. We apply the model to screening mammography interpretation: predicting the presence or absence of benign and malignant lesions. On the NYU Breast Cancer Screening Dataset, our model outperforms (AUC = 0.93) ResNet-34 and Faster R-CNN in classifying breasts with malignant findings. On the CBIS-DDSM dataset, our model achieves performance (AUC = 0.858) on par with state-of-the-art approaches. Compared to ResNet-34, our model is 4.1x faster for inference while using 78.4% less GPU memory. Furthermore, we demonstrate, in a reader study, that our model surpasses radiologist-level AUC by a margin of 0.11.

In this study, we investigate the feasibility of using dynamic contrast enhanced magnetic resonance imaging (DCE-MRI), diffusion weighted imaging (DWI), and dynamic positron emission tomography (PET) for detection of metastatic lymph nodes in head and neck squamous cell carcinoma (HNSCC) cases. Twenty HNSCC patients scheduled for lymph node dissection underwent DCE-MRI, dynamic PET, and DWI using a PET-MR scanner within one week prior to their planned surgery. During surgery, resected nodes were labeled to identify their nodal levels and sent for routine clinical pathology evaluation. Quantitative parameters of metastatic and normal nodes were calculated from DCE-MRI (v_e, v_p, PS, F_p, K^trans), DWI (ADC) and PET (K_i, K₁, k₂, k₃) to assess if an individual or a combination of parameters can classify normal and metastatic lymph nodes accurately. There were 38 normal and 11 metastatic nodes covered by all three imaging methods and confirmed by pathology. 34% of all normal nodes had volumes greater than or equal to the smallest metastatic node while 4 normal nodes had SUV > 4.5. Among the MRI parameters, the median v_p, F_p, PS, and K^trans values of the metastatic lymph nodes were significantly lower (p = <0.05) than those of normal nodes. v_e and ADC did not show any statistical significance. For the dynamic PET parameters, the metastatic nodes had significantly higher k₃ (p value = 8.8 × 10^-8) and K_i (p value = 5.3 × 10^-8) than normal nodes. K₁ and k₂ did not show any statistically significant difference. K_i had the best separation with accuracy = 0.96 (sensitivity = 1, specificity = 0.95) using a cutoff of K_i = 5.3 × 10^-3 mL/cm³/min, while k₃ and volume had accuracy of 0.94 (sensitivity = 0.82, specificity = 0.97) and 0.90 (sensitivity = 0.64, specificity = 0.97) respectively. 100% accuracy can be achieved using a multivariate logistic regression model of MRI parameters after thresholding the data with K_i < 5.3 × 10^-3 mL/cm³/min. The results of this preliminary study suggest that quantitative MRI may provide additional value in distinguishing metastatic nodes, particularly among small nodes, when used together with FDG-PET.

RATIONALE AND OBJECTIVES/OBJECTIVE:Pathologic complete response (pCR) in patients with human epidermal growth factor receptor 2 (HER2)-positive breast cancer after HER2-targeted therapy correlates increased disease-free survival and decreased mastectomy rates. The aim of this study was to explore tumor shrinkage patterns and initial tumor enhancement with pCR in HER2-positive breast cancer. MATERIALS AND METHODS/METHODS:This was an institutional review board-approved retrospective analysis of 51 HER2 positive breast cancer patients with breast MRI both pre- and post-HER2-targeted therapy. Initial enhancement ratio (IER, initial enhancement percentage over baseline at first postcontrast imaging), pattern of tumor shrinkage, and Dynamic contrast enhanced (DCE)-MRI imaging features were assessed. Wilcoxon rank, Spearman correlation, Fisher's exact, and Mann-Whitney tests were used to correlate MRI imaging features with pCR. IER reader agreement was evaluated by intraclass correlation. Binary logistic regression was used to evaluate multivariate associations with pCR. RESULTS:56.9% (29/51) of patients had pCR at surgery. Concentric tumor shrinkage pattern was associated with pCR (p = 0.001, Area under the curve (AUC) 0.778): accuracy 80.4%, specificity 96.6%, and sensitivity of 59.1%. There was no association with pCR and imaging response as defined by RECIST criteria (p = 0.169), pretreatment IER (Reader 1 (R1) p = 0.665, Reader 2 (R2) p = 0.766), or lesion size (p = 0.69). IER was associated with axillary metastases (R1 p = 0.016, R2 < 0.001) and ki-67 (R1 r = 0.52, p = 0.008, R2 r = -0.44, p = 0.028). CONCLUSION/CONCLUSIONS:The shrinkage pattern of HER2-positive tumors after targeted therapy may be associated with pCR. There was no association between IER and pCR. Future studies evaluating the correlation of shrinkage patterns to texture radiomics are of interest.

We present a deep convolutional neural network for breast cancer screening exam classification, trained and evaluated on over 200,000 exams (over 1,000,000 images). Our network achieves an AUC of 0.895 in predicting the presence of cancer in the breast, when tested on the screening population. We attribute the high accuracy to a few technical advances. (i) Our network's novel two-stage architecture and training procedure, which allows us to use a high-capacity patch-level network to learn from pixel-level labels alongside a network learning from macroscopic breast-level labels. (ii) A custom ResNet-based network used as a building block of our model, whose balance of depth and width is optimized for high-resolution medical images. (iii) Pretraining the network on screening BI-RADS classification, a related task with more noisy labels. (iv) Combining multiple input views in an optimal way among a number of possible choices. To validate our model, we conducted a reader study with 14 readers, each reading 720 screening mammogram exams, and show that our model is as accurate as experienced radiologists when presented with the same data. We also show that a hybrid model, averaging the probability of malignancy predicted by a radiologist with a prediction of our neural network, is more accurate than either of the two separately. To further understand our results, we conduct a thorough analysis of our network's performance on different subpopulations of the screening population, the model's design, training procedure, errors, and properties of its internal representations. Our best models are publicly available at https://github.com/nyukat/breastcancerclassifier.