Deep Learning Substitutes Gadolinium in Detecting Functional and Structural Brain Lesions with MRI

†These authors contributed equally and are joint first authors. 1Department of Electrical Engineering and the Taub Institute, Columbia University, New York, NY, USA 2Department of Biological Sciences and the Taub Institute, Columbia University, New York, NY, USA 3Department of Biomedical Engineering, Columbia University, New York, NY, USA 4Department of Radiation Oncology and the Irving Medical Center, Columbia University, New York, NY, USA 5Department of Pathology and Cell Biology, and the Irving Medical Center, Columbia University, New York, NY, USA 6Department of Radiation Oncology, the Irving Medical Center and the Herbert Irving Comprehensive Cancer Center, Columbia University, New York, NY, USA 7Department of Psychiatry, Columbia University, and the New York State Psychiatric Institute, New York, NY, USA 8Department of Neurology and the Taub Institute, Columbia University, New York, NY, USA 9Department of Neurology, the Taub Institute, the Sergievsky Center, Radiology and Psychiatry, Columbia University, New York, NY, USA 10Department of Psychiatry, Mortimer B. Zuckerman Mind Brain Behavior Institute, Columbia University, New York, NY, USA Email: jg3400@columbia.edu 11Data used in preparation of this article were partially obtained from the Alzheimer’s Disease Neuroimaging Initiative (ADNI) database (adni.loni.usc.edu). As such, the investigators within the ADNI contributed to the design and implementation of ADNI and/or provided data but did not participate in analysis or writing of this report. A complete listing of ADNI investigators can be found at: http://adni.loni.usc.edu/wp-content/uploads/how_to_apply/ADNI_ Acknowledgement_List.pdf

: Overview of the studies conducted. We first performed proof-of-concept studies in mice to validate our hypothesis that deep learning can extract information equivalent to Gadolinium-based contrast agent (GBCA) contrast enhancement from a single-modal non-contrast MRI scan, and then conducted extensive analyses in humans to scrutinize the capability of this proposed approach. Study: study conducted; Aim: purpose of the study; Source: where the imaging data come from; MRI: modality/type of data used in the study; Model: specific DeepContrast model used in the study; Outcome: specific utility of GBCA replicated by DeepContrast. AD: Alzheimer's disease; ADNI: Alzheimer's Disease Neuroimaging Initiative dataset; BraTS: Brain Tumor Segmentation dataset; CBV: cerebral blood volume; CHR: clinical high-risk for Schizophrenia; Gd-Uptake: GBCA contrast uptake maps; GBM: glioblastoma multiforme; T2W: T2-weighted scans; T2W-CE: T2-weighted contrast-enhanced scans; T1W: T1-weighted scans; T1W-CE: T1-weighted contrastenhanced scans.
we demonstrated that the method can be generalized to other organs, by showing that it can enhance structural lesions caused by breast tumors. The deep learning model will be referred to as 'DeepContrast'. 53 The studies conducted are outlined in Fig. 1. 54

55
DeepContrast in the mouse brain 56 We first designed, optimized and trained the model on wildtype (WT) mice brain scans (37 for training 57 and 6 for validation; see methods), in which we have previously generated quantitative T2-weighted 58 GBCA-uptake brain maps. Similar to previous studies 14, 15 , we compared the similarities between the 59 GBCA-predicted maps and the GBCA-uptake ground truth maps by performing voxel-wise analyses 60 across the whole brain on a test set with 6 scans (Fig. 2b) using metrics that measure signal quality (peak 61 Fig. 2: Quantitative evaluation of the DeepContrast in the mouse brain. a. DeepContrast prediction (Gd-Predicted) highly concords with the ground truth GBCA-uptake map (Gd-Uptake) in the mouse brain. The non-contrast scans and the contrast-enhanced scans are displayed for reference. A healthy wild type (WT) subject is shown in the left panel while a subject with glioblastoma multiforme (GBM) is shown on the right panel. Color bars indicate the colormap and dynamic range used in the crosssectional brain images. b-e. Similarity between the model prediction and the ground truth, evaluated on all scans in the test sets (b: WT, 6 scans; c-e: GBM, 4 scans) using quantitative metrics, where the non-contrast (T2W) scans are used as the performance baseline. f. ROC curves for mouse GBM high-enhancement region similarity assessment on the test set: ROC curve for the model prediction in comparison to the ground truth GBCA-uptake map (left) and ROC curve for the non-contrast (T2W) scans in comparison to the ground truth GBCA-uptake map (right). ROC curve for the model prediction (sensitivity = 0.72, specificity = 0.88, AUC = 0.85) outperforms the ROC curve for the non-contrast (T2W) scans (sensitivity = 0.50, specificity = 0.75, AUC = 0.64). The standard deviation is indicated by the shaded area. For all voxel-based metrics, only the voxels within the brains or subregions are used. SSIM is calculated on the minimum bounding box of the brains or subregions. Asterisks indicate level of statistical significance (* p < 0.05, ** p < 0.01, *** p < 0.001, *** p < 0.0001). PSNR: peak signal-to-noise ratio; SSIM: structural similarity index; P.R: Pearson correlation coefficient; S.R: Spearman correlation coefficient. DeepContrast in the human brain 80 We adapted the DeepContrast model to human brain MRI datasets by modifying the network architecture, 81 hyper-parameters and training strategies. First, same as the mouse study, we compared the similarities 82 between the GBCA-predicted maps and the GBCA-uptake ground truth maps by performing voxel-wise 83 analyses across the whole brain on a test set with 179 scans (Fig. 3a-b). Between the maps, the peak 84 Fig. 3: Quantitative evaluation of the DeepContrast in the cognitively normal human brain. a. DeepContrast prediction (Gd-Predicted) highly concords with the ground truth GBCA-uptake map (Gd-Uptake) in the cognitive normal human brain. Color bars indicate the colormap and dynamic range used in the cross-sectional brain images. b. Similarity between the model prediction and the ground truth, evaluated on 179 scans of cognitively normal (CN) subjects using quantitative metrics, where non-contrast scans are used as the performance baseline. c. DeepContrast shows higher test-retest reliability than the experimentally acquired Gd-Uptake ground truth. For all voxel-based metrics, only the voxels within the brains or subregions are used. SSIM is calculated on the minimum bounding box of the brains or subregions. Asterisks indicate level of statistical significance (* p < 0.05, ** p < 0.01, *** p < 0.001, **** p < 0.0001). PSNR: peak signal-to-noise ratio; SSIM: structural similarity index; P.R: Pearson correlation coefficient; S.R: Spearman correlation coefficient. signal-to-noise ratio was 29.64±0.07, the Pearson correlation coefficient was 0.822±0.002 (p < 0.0001), Fig. 4: DeepContrast maps differential anatomical patterns of dysfunction in the hippocampal formation. a. A threedimensional rendering of the bilateral hippocampal formation (left panel) consisting of the hippocampus (HC) and the entorhinal cortex (EC) and axial, sagittal, and coronal slices from a group-wise T1-weighted MRI template cutting through the hippocampal formation (right three panels). The hippocampal formation is displayed with the hot-to-cold colormap along the anterior-toposterior axis. b. A voxel-based analysis on the CBV-predicted maps of 177 individuals ranging from 20-72 years of age reveals that the greatest age-related decline occurred in the body of the hippocampal circuit (top, color-coded by the degree of significance). A coronal slice (bottom), onto which the hippocampal formation mask is applied, reveals that age-related decline localizes primarily to the dentate gyrus. The voxel-based analysis is conducted using a multiple regression model in SPM12 using sex as a covariate and age as the regressor, and the age-related differences are contrasted using Student's t test. Multiple comparisons are corrected for, yielding voxel-wise p < 0.005 and cluster-wise p < 0.05 (see methods). c. A voxel-based analysis on the CBV-predicted maps of 74 Schizophrenia clinical high-risk (CHR) patients with 18 normal controls reveals CHR-related increase in the body of the hippocampal circuit (top, color-coded by the degree of significance). A coronal slice (bottom), onto which the hippocampal formation mask is applied, reveals that CHR-related increase localizes primarily to the CA1. The voxel-based analysis is conducted using a general linear model in SPM12 using a two-sample student's t test after controlling for global variables, and the CHRrelated differences are contrasted using Student's t test. Multiple comparisons are corrected for, yielding voxel-wise p < 0.005 and cluster-wise p = 0.3 (see methods). d. A voxel-based analysis on the CBV-predicted maps of 50 Alzheimer's disease (AD) patients compared with 50 normal controls, each with 2 back-to-back scans, reveals AD-related reduction in the entorhinal cortex (top, color-coded by the degree of significance). A coronal slice (bottom), onto which the hippocampal formation mask is applied, reveals that AD-related decline localizes primarily to the transentorhinal cortex. The voxel-based analysis is conducted using a multiple regression model in SPM12 using age, sex and subject identity as covariates and diagnostic class (i.e., cognitive normal vs. dementia) as the regressor, and the AD-related difference are contrasted using Student's t test. Multiple comparisons are corrected for, yielding voxel-wise p < 0.005 and cluster-wise p < 0.05 (see methods). e. A scatter plot shows the association between age and mean CBV-Predicted values in the dentate gyrus after removal of gender effects (βage = −6.36e-4, tage = −4.64, page = 6.85e-6  were significantly lower (p = 0.031) in AD patients compared to the healthy controls (Fig. 4g).

146
DeepContrast enhances structural lesions 147 Brain Tumor. In order to accurately capture the high variance present within brain tumors, we retrained 148 a DeepContrast model with a large-scale brain tumor MRI dataset collected from the Brain Tumor 149 Segmentation (BraTS) [39][40][41][42] database. First, we compared the similarities between the GBCA-predicted 150 maps and the GBCA-uptake ground truth maps by performing voxel-wise analyses across the whole 151 brain on a test set with 15 scans (Fig. 6a-b). Between the maps, the peak signal-to-noise ratio was   Table 1. Finally, we conducted the ROC analysis, and the average ROC curve reached a 156 sensitivity of 0.81 and a specificity of 0.85 at the operating point, whereas the AUC was 0.91 (Fig. 6c).
Breast Tumor. We further explored the possibility of extending DeepContrast to other organs. We test set with 16 scans (Fig. 6d-e). Between the maps, the peak signal-to-noise ratio was 27.40±0.64, the 166 Pearson correlation coefficient was 0.691±0.034 (p < 0.0001), the Spearman correlation coefficient was 0.630±0.023 (p < 0.0001), and the structural similarity index was 0.826±0.021. Next, we calculated the 168 same metrics in tumor and non-tumor regions, and the results were illustrated in Fig. 6e and reported in 169 Table 1. Finally, we conducted the ROC analysis, and the average ROC curve reached a sensitivity of 170 0.77 and a specificity of 0.82 at the operating point, whereas the AUC was 0.87 (Fig. 6f). 171 Fig. 6: DeepContrast enhances structural lesions in human brain and breast MRIs. a. DeepContrast prediction (Gd-Predicted) highly concords with the ground truth GBCA-uptake map (Gd-Uptake) of structural lesions in human brain. Color bars indicate the colormap and dynamic range used in the cross-sectional brain images. b. Similarity between the model prediction and the ground truth, evaluated on 15 scans of subjects with glioblastoma multiforme (GBM) using quantitative metrics, where non-contrast scans are used as the performance baseline. c. ROC curves for human GBM high-enhancement region similarity assessment on the test set: ROC curve for the model prediction in comparison to the ground truth GBCA-uptake map (top) and ROC curve for the non-contrast (T1W) scans in comparison to the ground truth GBCA-uptake map (bottom). ROC curve for the model prediction (sensitivity = 0.81, specificity = 0.85, AUC = 0.91) outperforms the ROC curve for the non-contrast (T1W) scans (sensitivity = 0.97, specificity = 0.07, AUC = 0.25). The standard deviation is indicated by the shaded area. d. DeepContrast prediction (Gd-Predicted) highly concords with the ground truth GBCA-uptake map (Gd-Uptake) of structural lesions in human breast. Color bars indicate the colormap and dynamic range used in the cross-sectional breast images. e. Similarity between the model prediction and the ground truth, evaluated on 16 scans of subjects with breast tumor using quantitative metrics, where non-contrast scans are used as the performance baseline. f. ROC curves for breast tumor high-enhancement region similarity assessment on the test set: ROC curve for the model prediction in comparison to the ground truth GBCA-uptake map (left) and ROC curve for the non-contrast (T1W) scans in comparison to the ground truth GBCA-uptake map (right). ROC curve for the model prediction (sensitivity = 0.77, specificity = 0.82, AUC = 0.87) outperforms the ROC curve for the non-contrast (T1W) scans (sensitivity = 0.59, specificity = 0.70, AUC = 0.70). The standard deviation is indicated by the shaded area. For all voxel-based metrics, only the voxels within the brains, breasts or subregions are used. SSIM is calculated on the minimum bounding box of the brains, breasts or subregions. Asterisks indicate level of statistical significance (* p < 0.05, ** p < 0.01, *** p < 0.001, **** p < 0.0001). PSNR: peak signal-to-noise ratio; SSIM: structural similarity index; P.R: Pearson correlation coefficient; S.R: Spearman correlation coefficient. Evaluations varied depending on the aspects being assessed for each model. All metrics reported in the form of mean ± standard error of the mean (SEM). PSNR: peak signal-to-noise ratio; P.R: Pearson correlation coefficient; S.R: Spearman correlation coefficient; SSIM: structural similarity index.

172
By using a quantitative GBCA dataset in mice and humans, we demonstrated that deep learning can, in 173 principle, generate GBCA-equivalent information from a single and common MRI scan across an array of 174 lesions.

175
GBCA's utility for MRI can be organized around two primary pathophysiologies. The first is a 176 breakdown of the blood-brain barrier that often accompanies many structural lesions, and in which case 177 GBCA extravasates into the parenchyma and enhances lesion detection 47 . The second is alterations in 178 neuronal metabolism, typifying most functional disorders, in which case intravascular GBCA is used 179 to quantify regional CBV, a hemodynamic variable tightly coupled to energy metabolism 4, 48-51 . As 180 proof-of-principle, we optimized five models for our investigations across two species, two organs and 181 multiple disorders. As GBCA's utility can be reduced to two pathophysiologies we anticipate that 182 future large-scale studies across a range of diseases might lead to two generalizable models-one for 183 structural disorders that break down the blood-brain barrier, another for functional lesions that alters 184 brain metabolism.  DeepContrast's utility can be organized according to its broad applications. The first is for research.
There is an increasing number of brain MRI databases, such as ADNI (see Supplementary Table 2 for 193 an example list of more than twenty open datasets), whose sole purpose is brain imaging and disease 194 research. Standard T1-weighted MRI scans are among the most common acquisition across all of these 195 datasets, typically acquired for mapping regional structural differences, such as regional volume or cortical 196 thickness. DeepContrast can be retroactively applied to these datasets, and can be used to generate          Healthy Mouse Brain. In total, 49 WT mice were used in this study. Whole brain T2W MRI scans 282 before (T2W) and 35 mins after intraperitoneal injection (T2W-CE) of Gadodiamide at 10 mmol/kg were 283 acquired with identical scan parameters as previously described in CBV-fMRI protocol. The Gd-Uptake 284 ground truth was quantified with the standardized delta-R2, which was derived using the same method 285 as discussed before 20 , followed by a standardization to the dynamic range of [0, 1]. We used 3D PCNN 54 286 with manual correction to generate brain masks, which we used as training fields over which the model 287 was optimized and performance metrics were calculated. A train-validation-test ratio at 8:1:1 was applied 288 in the healthy mouse brain model training. Mouse GBM. For scans of tumor subjects, the CBV maps and brain masks were derived using the 291 same methods as described in the healthy mouse brain study, and tumor masks were generated in 292 addition to the brain masks using the Fuzzy-C-Means segmentation 55 . 6 GBM subjects were added  Healthy Human Brain. T1-weighted MRI scans were acquired using the protocols as described 296 previously 17, 18 , before (T1W) and 4 minutes after (T1W-CE) intravenous injection of Gadodiamide.

297
During the MRI acquisition for the same session, the receiver gain was kept constant and the offset 298 was set to zero, and as a result the T1W and T1W-CE scans share the same scaling and zero shifting.

299
Each T1W & T1W-CE pair was spatially aligned when provided. For intensity normalization, each 300 T1W scan was compressed to the dynamic range of [0, 1], and the corresponding T1W-CE scan was scaled accordingly to match the constant scaling. The Gd-Uptake ground truth was quantified with the 302 steady-state MRI method 17 , by subtracting the normalized T1W scans from the respective T1W-CE 303 scans. We generated brain masks using the BET function in FMRIB Software Library (FSL) 56 , which we 304 used as training fields over which the model was optimized and performance metrics were calculated. 305 We generated tissue label maps using the FAST function in FSL for tissue-of-interest analyses. The   Human CHR. The 92-scan cohort for the CHR study was acquired using the same scan parameters as 316 those used to train the DeepContrast Healthy Human Brain Model, which ensures minimal discrepancy 317 in scan appearance. CHR patients present very little structural deformation, which ensures minimal 318 discrepancy in scan anatomy. Therefore, no additional measure needs to be taken to deal with 319 appearance or anatomy variances. After normalization to the dynamic range of [0, 1], the scans 320 were directly treated as inputs to the model to generate the Gd-Predicted maps. CBV-predicted 321 maps were then generated by applying the same normalization method as we would to quantify CBV maps. thus resulting in the mismatch in anatomy. We approached these issues by first minimizing the 332 between-cohort appearance difference using a dynamic histogram warping (DHW) algorithm 57 as it was 333 demonstrated to be among the best intensity matching methods in medical imaging 58 . Specifically, we 334 calculated the mean normalized-brain-region 2048-bin histogram of each cohort, derived a bin-to-bin 335 mapping between the cohorts, and applied the mapping to each individual scan in the AD study. 336 Secondly, we minimized the anatomical difference by running a diffeomorphic registration 68 prior to 337 applying the DeepContrast model. After these two steps, we normalized the scans to the dynamic 338 range of [0, 1] and provided them to the model to generate the Gd-Predicted maps. CBV-predicted 339 maps were then generated by applying the same normalization method as we would to quantify CBV maps. spatial matching with the T1W and T1W-CE pairs. Whole breast masks were generated using k-means 359 clustering. T1W and T1W-CE pairs were both normalized using the maximum of the T1W scans before 360 being fed into the DeepContrast model.

361
DeepContrast and its implementations 362 All five model variants developed in our studies, as mentioned in Fig. 1, shared the common residual 363 attention U-Net (RAU-Net) architecture (Fig. 7). Model inputs were the non-contrast MRI scans, while 364 the outputs were the corresponding predicted GBCA contrast (Gd-Predicted). The inputs and outputs 365 were in equal dimensions and were either 2D or 3D depending on the nature of the scan protocols (i.e.,

366
2D slices were used for 2D MRI scans, whereas 3D volumes were used for 3D MRI scans). Preprocessing includes intensity normalization and brain extraction. Ground truth Gd-Uptake was derived using the standardized delta-R2 equation. Note that there is an additional standardization step that maps the dynamic range of the standardized delta-R2 to the range of [0, 1], before the application of the brain mask. The loss function is calculated between the Gd-Uptake and the predicted version only using the voxels within the brain mask region. b. This training strategy applies to the Healthy Human Brain Model. Preprocessing includes intensity normalization and brain extraction. Ground truth Gd-Uptake was derived using the steady-state delta-R1 equation. The loss function is calculated between the Gd-Uptake and the predicted version only using the voxels within the brain mask region. c. This training strategy applies to the Tumor Human Brain Model. Preprocessing includes intensity normalization and brain extraction. Ground truth Gd-Uptake was derived using the steady-state delta-R1 equation. The loss function is calculated between the Gd-Uptake and the predicted version only using the voxels within the brain mask region. d. This training strategy applies to the Tumor Human Breast Model. Preprocessing includes intensity normalization and brain extraction. Ground truth Gd-Uptake was derived using the steady-state delta-R1 equation. The loss function is calculated between the Gd-Uptake and the predicted version only using the voxels within the breast mask region.
image space information. The add-on residual blocks simplify the entities to be approximated across each layer and therefore enables training of deeper networks, while the attention gates learn to differentially 375 enhance or suppress specific regions in the feature maps so that the downstream outcomes are better 376 represented for targeting objective.   Fig. 4) 388 was a 2D RAU-Net that consisted of 5 encoding and decoding layers. The model input was a 2D axial 389 slice of the mouse brain scans. Adam optimizer with a learning rate of 0.001 was used in this study. Our   Tumor Human Brain Model. The model used in the human tumor brain study (Supplementary Fig. 6) 404 was a 3D RAU-Net that consisted of 6 encoding and decoding layers. The model input was a 3D human 405 brain volume. SGD optimizer with an adaptive learning rate handle with 0.001 initial learning rate was 406 used in this study. Our batch size was 1 and the robust adaptive loss function was utilized. TorchIO 65 was 407 used for data augmentation to counteract the insufficiency of tumor variance presented in the training data.  This was done using an Otsu filter 66 which automatically selected the threshold value dividing the voxels 431 into 2 classes. The prediction candidate was binarized for 1,000 times using 1,000 evenly distributed 432 thresholds between the minimum and maximum of the candidate. The ROC curve was then created by 433 comparing the 1,000 versions of the prediction candidate to the binarized ground truth using Scikit-learn 67 .

435
Voxel-based analysis for regional vulnerability localization: Human Aging. Voxel-based 436 analysis (Fig. 4b) was performed by first transforming the non-contrast images using a diffeomorphic 437 registration algorithm 68 with nearest-neighbor interpolation to an unbiased brain template created from 438 the 177 scans in the Aging study 68 . The GBCA-predicted maps were generated by the Healthy Human 439 Brain model using the native-space non-contrast T1W scans as the input and were subsequently used to 440 generate CBV-predicted maps by normalization using mean value among the top 10% brightest voxels 441 within the brain region (representing signal intensity from pure blood). These CBV-predicted maps 442 were then transformed into the template using the same transformation parameters calculated from the 443 registration process, and subsequently smoothed using a 3mm-diameter spherical kernel. Transformed 444 and filtered CBV-predicted maps were analyzed using SPM12 69 . Data were analyzed with a multiple 445 regression model, including sex as a covariate and age as the regressor. Age-related differences were 446 contrasted using Student's t test. FreeSurfer regional segmentation was then performed on the unbiased 447 template image, and the hippocampal formation mask is generated by binarizing and combining the labels  Whole Brain Aging Analysis. The GBCA-predicted maps were generated in the native space 464 of each subject and were afterwards used for CBV quantification together with the experimentally 465 acquired ground truth GBCA-uptake maps using the same whole brain top 10% mean normalization. 466 Similarly, the T1W scans were normalized to generate a comparable counterpart. We used T1W 467 scans for comparison because they were the only input to the DeepContrast model to generate 468 GBCA-predicted maps. The CBV (quantified from Gd-Uptake), CBV-Predicted (quantified from 469 Gd-Predicted), and normalized T1W scans were used for age-related regression in the multiple brain  Voxel-based analysis for regional vulnerability localization: Human CHR. Voxel-based 485 analysis (Fig. 4c) was performed by first transforming the non-contrast images using a diffeomorphic 486 registration algorithm 68 with nearest-neighbor interpolation to an unbiased brain template created 487 from the 92 scans from the CHR study 68 . The GBCA-predicted maps were generated by the Healthy Slice-based analysis for regional vulnerability localization: Human CHR. Slice-based analysis 510 ( Supplementary Fig. 3) was performed by first transforming the non-contrast images using a diffeomorphic 511 registration algorithm 68 with nearest-neighbor interpolation to an unbiased brain template created from 512 the 92-scan population 68 . The GBCA-predicted map was generated by the Healthy Human Brain model 513 using the native-space non-contrast T1W scans as the input and was subsequently used to quantify 514 CBV-predicted maps by normalizing them by their respective mean value among the top 10% brightest 515 voxels within the brain region. These CBV-predicted maps were then transformed into that template 516 using the identical transformation parameters calculated from the registration process, and subsequently 517 smoothed using a 3mm-diameter spherical kernel. Next we upsampled the unbiased template as well as 518 these CBV-predicted scans to an isotropic resolution (voxel size = 0.68×0.68×0.68 mm) using cubic 519 spline interpolation. We parcellated the hippocampal subfields of the template using FreeSurfer, and 520 further cut the left and right CA1 subregions in the hippocampus into slices along the anterior-posterior 521 axes of these structures. We computed the slice-mean CBV-Predicted values for each slice, followed by a 522 3-slice sliding window averaging to smooth the results. Then we ran two-sample t-tests as discussed 523 before over the smoothed slice-mean CBV-Predicted values to generate the slice-based analysis results.

525
Voxel-based analysis for regional vulnerability localization: Human AD. Voxel-based analysis 526 (Fig. 4d) was performed by first transforming the non-contrast images using a diffeomorphic registration 527 algorithm 68 with nearest-neighbor interpolation to an unbiased brain template created from the 200 528 scans (i.e., 100 subjects each with 2 back-to-back repeated scans) in the AD study 68 . We then ran these 529 non-contrast scans through the DeepContrast Healthy Human Brain Model to generate CBV-predicted 530 maps, which were subsequently smoothed using a 3mm-diameter spherical kernel. Unlike in the aging 531 study, the application of DeepContrast was performed after the registration process to help eliminate 532 major anatomical variances, since the deformations present in the diseased population were not previously 533 observed by the model trained on healthy data. GBCA-predicted scans, the direct output of the model, 534 were used to quantify CBV-predicted maps using the same method as described in the Aging study above.

535
These CBV-predicted maps, already co-registered upon creation, were analyzed using SPM12. Data 536 were analyzed with a multiple regression model, including age, sex and subject identity as covariates 537 and diagnostic class (i.e., cognitive normal vs. dementia) as the regressor. AD-related differences were 538 contrasted using Student's t test. FreeSurfer regional segmentation was then performed on the unbiased 539 template image, and the hippocampal formation mask was generated by binarizing and combining the 540 labels corresponding to the hippocampus and the entorhinal cortex, while an extended hippocampal 541 formation mask was additionally generated to also include the parahippocampal cortex. The AD-related 542 regression t-map was then projected onto the MNI-152 brain template using diffeomorphic transformation 543 with nearest-neighbor interpolation. The result was thresholded at p < 0.005 and corrected for multiple 544 comparisons at the cluster level within the extended hippocampal formation using a Monte-Carlo 545 simulation implemented in AFNI-3dClustSim (10,000 iterations) to yield a corrected p < 0.05. The final corrected AD-related regression t-map was then overlaid onto the MNI-152 template in cross-section 547 using 3DSlicer, and also displayed with composite-with-shading volume rendering over semi-transparent 548 models of the hippocampal formation. were used to conduct the right transentorhinal cortex (TEC) ROI analysis. A two-sample t-test was 552 conducted over the right TEC, at the boundary between the right entorhinal cortex (EC) and the right 553 parahippocampal cortex (PHC). The region was defined as the intersection between the EC-PHC region 554 and a sphere centered at the middle of the EC-PHC intersection and spanning a diameter of the extent of 555 the EC-PHC boundary (11 mm). A box plot overlaid with individual data points was drawn (Fig. 4g) to 556 indicate the group-wise difference between the normal controls and the AD patients.

557
Data Availability

558
The trained Healthy Human Brain Model, alongside the test-retest reliability dataset (n = 11, each 559 with two test-retest acquisitions) with both non-contrast scans and ground truth GBCA-uptake maps, 560 is available on GitHub (link to be announced). The scripts that predict GBCA-uptake maps from commented and edited the manuscript. Data used in the Human AD study was obtained from the ADNI 618 database (adni.loni.usc.edu). As such, the investigators within the ADNI contributed to the design 619 and implementation of ADNI and/or provided data but did not participate in analysis or writing of this 620 manuscript.  Fig. 7). This model is implemented with a 2D six-layered Residual Attention U-Net architecture. The encoding path (blue blocks in the left half of the architecture) condenses the image dimensions and enriches the feature dimension, shrinking the image size from 352×352 pixels to 11×11 pixels while extracting 2048 channels of features. The decoding path (red blocks in the right half of the architecture) expands these high-level features and returns back a single slice of the predicted image with 352×352 pixels.
Supplementary Fig. 6: Details of the tumor human brain model (Network 3 in Fig. 7). This model is implemented with a 3D six-layered Residual Attention U-Net architecture. The encoding path (blue blocks in the left half of the architecture) condenses the image dimensions and enriches the feature dimension, shrinking the image size from 192×192×160 pixels to 6×6×5 pixels while extracting 256 channels of features. The decoding path (red blocks in the right half of the architecture) expands these high-level features and returns back a single slice of the predicted image with 192×192×160 pixels.
Supplementary Fig. 7: Details of the tumor human breast model (Network 4 in Fig. 7). This model is implemented with a 2D five-layered Residual Attention U-Net architecture. The encoding path (blue blocks in the left half of the architecture) condenses the image dimensions and enriches the feature dimension, shrinking the image size from 256×256 pixels to 16×16 pixels while extracting 256 channels of features. The decoding path (red blocks in the right half of the architecture) expands these high-level features and returns back a single slice of the predicted image with 256×256 pixels.