Counterfactual MRI Generation with Denoising Diffusion Models for Interpretable Alzheimer's Disease Effect Detection.

Generative AI models have recently achieved mainstream attention with the advent of powerful approaches such as stable diffusion, DALL-E and MidJourney. The underlying breakthrough generative mechanism of denoising diffusion modeling can generate high quality synthetic images and can learn the underlying distribution of complex, high-dimensional data. Recent research has begun to extend these models to medical and specifically neuroimaging data. Typical neuroimaging tasks such as diagnostic classification and predictive modeling often rely on deep learning approaches based on convolutional neural networks (CNNs) and vision transformers (ViTs), with additional steps to help in interpreting the results. In our paper, we train conditional latent diffusion models (LDM) and denoising diffusion probabilistic models (DDPM) to provide insight into Alzheimer's disease (AD) effects on the brain's anatomy at the individual level. We first created diffusion models that could generate synthetic MRIs, by training them on real 3D T1-weighted MRI scans, and conditioning the generative process on the clinical diagnosis as a context variable. We conducted experiments to overcome limitations in training dataset size, compute time and memory resources, testing different model sizes, effects of pretraining, training duration, and latent diffusion models. We tested the sampling quality of the disease-conditioned diffusion using metrics to assess realism and diversity of the generated synthetic MRIs. We also evaluated the ability of diffusion models to conditionally sample MRI brains using a 3D CNN-based disease classifier relative to real MRIs. In our experiments, the diffusion models generated synthetic data that helped to train an AD classifier (using only 500 real training scans) -and boosted its performance by over 3% when tested on real MRI scans. Further, we used implicit classifier-free guidance to alter the conditioning of an encoded individual scan to its counterfactual (representing a healthy subject of the same age and sex) while preserving subject-specific image details. From this counterfactual image (where the same person appears healthy), a personalized disease map was generated to identify possible disease effects on the brain. Our approach efficiently generates realistic and diverse synthetic data, and may create interpretable AI-based maps for neuroscience research and clinical diagnostic applications.

Few-Shot Classification of Autism Spectrum Disorder using Site-Agnostic Meta-Learning and Brain MRI.

For machine learning applications in medical imaging, the availability of training data is often limited, which hampers the design of radiological classifiers for subtle conditions such as autism spectrum disorder (ASD). Transfer learning is one method to counter this problem of low training data regimes. Here we explore the use of meta-learning for very low data regimes in the context of having prior data from multiple sites - an approach we term site-agnostic meta-learning. Inspired by the effectiveness of meta-learning for optimizing a model across multiple tasks, here we propose a framework to adapt it to learn across multiple sites. We tested our meta-learning model for classifying ASD versus typically developing controls in 2,201 T1-weighted (T1-w) MRI scans collected from 38 imaging sites as part of Autism Brain Imaging Data Exchange (ABIDE) [age: 5.2 -64.0 years]. The method was trained to find a good initialization state for our model that can quickly adapt to data from new unseen sites by fine-tuning on the limited data that is available. The proposed method achieved an area under the receiver operating characteristic curve (ROC-AUC)=0.857 on 370 scans from 7 unseen sites in ABIDE using a few-shot setting of 2-way 20-shot i.e., 20 training samples per site. Our results outperformed a transfer learning baseline by generalizing across a wider range of sites as well as other related prior work. We also tested our model in a zero-shot setting on an independent test site without any additional fine-tuning. Our experiments show the promise of the proposed site-agnostic meta-learning framework for challenging neuroimaging tasks involving multi-site heterogeneity with limited availability of training data.Clinical Relevance- We propose a learning framework that accommodates multi-site heterogeneity and limited data to assist in challenging neuroimaging tasks.

Along-Tract Statistical Mapping of Microstructural Abnormalities in Bipolar Disorder: A Pilot Study.

Investigating brain circuitry involved in bipolar disorder (BD) is key to discovering brain biomarkers for genetic and interventional studies of the disorder. Even so, prior research has not provided a fine-scale spatial mapping of brain microstructural differences in BD. In this pilot diffusion MRI dataset, we used BUndle ANalytics (BUAN)-a recently developed analytic approach for tractography-to extract, map, and visualize the profile of microstructural abnormalities on a 3D model of fiber tracts in people with BD (N=38) and healthy controls (N=49), and investigate along-tract white matter (WM) microstructural differences between these groups. Using the BUAN pipeline, BD was associated with lower mean fractional anisotropy (FA) in fronto-limbic and interhemispheric pathways and higher mean FA in posterior bundles relative to controls.Clinical Relevance- BUAN combines tractography and anatomical information to capture distinct along-tract effects on WM microstructure that may aid in classifying diseases based on anatomical differences.

Efficiently Training Vision Transformers on Structural MRI Scans for Alzheimer's Disease Detection.

Neuroimaging of large populations is valuable to identify factors that promote or resist brain disease, and to assist diagnosis, subtyping, and prognosis. Data-driven models such as convolutional neural networks (CNNs) have increasingly been applied to brain images to perform diagnostic and prognostic tasks by learning robust features. Vision transformers (ViT) - a new class of deep learning architectures - have emerged in recent years as an alternative to CNNs for several computer vision applications. Here we tested variants of the ViT architecture for a range of desired neuroimaging downstream tasks based on difficulty, in this case for sex and Alzheimer's disease (AD) classification based on 3D brain MRI. In our experiments, two vision transformer architecture variants achieved an AUC of 0.987 for sex and 0.892 for AD classification, respectively. We independently evaluated our models on data from two benchmark AD datasets. We achieved a performance boost of 5% and 9-10% upon fine-tuning vision transformer models pre-trained on synthetic (generated by a latent diffusion model) and real MRI scans, respectively. Our main contributions include testing the effects of different ViT training strategies including pre-training, data augmentation and learning rate warm-ups followed by annealing, as pertaining to the neuroimaging domain. These techniques are essential for training ViT-like models for neuroimaging applications where training data is usually limited. We also analyzed the effect of the amount of training data utilized on the test-time performance of the ViT via data-model scaling curves.Clinical Relevance- The models evaluated in this work could be trained on neuroimaging data to assist in diagnosis, subtyping and prognosis of Alzheimer's disease.

Evaluating Fiber Orientation Dispersion Measures Computed From Single-Shell Diffusion MRI.

Fiber orientation dispersion is one of the fundamental features that can be estimated from diffusion magnetic resonance imaging (dMRI) of the brain. Several approaches have been proposed to estimate dispersion from single- and multi-shell dMRI acquisitions. Here, we derive solutions to bring these proposed methods to a standard orientation dispersion index (ODI) with the goal of making them comparable across different dMRI acquisitions. To illustrate the utility of the measures in studying brain aging, we further examined the age-dependent trajectory of the different single- and multi-shell ODI estimates in the white matter across the lifespan.Clinical Relevance- This work computes metrics of brain microstructure that can be adapted for large neuroimaging initiatives that aim to study the brain's development and aging, and to identify deviations that may serve as biomarkers of brain disease.

Curriculum Based Multi-Task Learning for Parkinson's Disease Detection.

There is great interest in developing radiological classifiers for diagnosis, staging, and predictive modeling in progressive diseases such as Parkinson's disease (PD), a neurodegenerative disease that is difficult to detect in its early stages. Here we leverage severity-based meta-data on the stages of disease to define a curriculum for training a deep convolutional neural network (CNN). Typically, deep learning networks are trained by randomly selecting samples in each mini-batch. By contrast, curriculum learning is a training strategy that aims to boost classifier performance by starting with examples that are easier to classify. Here we define a curriculum to progressively increase the difficulty of the training data corresponding to the Hoehn and Yahr (H&Y) staging system for PD (total N=1,012; 653 PD patients, 359 controls; age range: 20.0-84.9 years). Even with our multi-task setting using pre-trained CNNs and transfer learning, PD classification based on T1-weighted (T1-w) MRI was challenging (ROC AUC: 0.59-0.65), but curriculum training boosted performance (by 3.9%) compared to our baseline model. Future work with multimodal imaging may further boost performance.

Transferring Models Trained on Natural Images to 3D MRI via Position Encoded Slice Models.

Transfer learning has remarkably improved computer vision. These advances also promise improvements in neuroimaging, where training set sizes are often small. However, various difficulties arise in directly applying models pretrained on natural images to radiologic images, such as MRIs. In particular, a mismatch in the input space (2D images vs. 3D MRIs) restricts the direct transfer of models, often forcing us to consider only a few MRI slices as input. To this end, we leverage the 2D-Slice-CNN architecture of Gupta et al. (2021), which embeds all the MRI slices with 2D encoders (neural networks that take 2D image input) and combines them via permutation-invariant layers. With the insight that the pretrained model can serve as the 2D encoder, we initialize the 2D encoder with ImageNet pretrained weights that outperform those initialized and trained from scratch on two neuroimaging tasks -- brain age prediction on the UK Biobank dataset and Alzheimer's disease detection on the ADNI dataset. Further, we improve the modeling capabilities of 2D-Slice models by incorporating spatial information through position embeddings, which can improve the performance in some cases.

Along-Tract Statistical Mapping of Microstructural Abnormalities in Bipolar Disorder: A Pilot Study.

Investigating brain circuitry involved in bipolar disorder (BD) is key to discovering brain biomarkers for genetic and interventional studies of the disorder. Even so, prior research has not provided a fine-scale spatial mapping of brain microstructural differences in BD. In this pilot diffusion MRI dataset, we used BUndle ANalytics (BUAN), a recently developed analytic approach for tractography, to extract, map, and visualize the profile of microstructural abnormalities on a 3D model of fiber tracts in people with BD (N=38) and healthy controls (N=49), and investigate along-tract white matter (WM) microstructural differences between these groups. Using the BUAN pipeline, BD was associated with lower mean Fractional Anisotropy (FA) in fronto-limbic and interhemispheric pathways and higher mean FA in posterior bundles relative to controls. BUAN combines tractography and anatomical information to capture distinct along-tract effects on WM microstructure that may aid in classifying diseases based on anatomical differences.

Efficiently Training Vision Transformers on Structural MRI Scans for Alzheimer's Disease Detection.

Neuroimaging of large populations is valuable to identify factors that promote or resist brain disease, and to assist diagnosis, subtyping, and prognosis. Data-driven models such as convolutional neural networks (CNNs) have increasingly been applied to brain images to perform diagnostic and prognostic tasks by learning robust features. Vision transformers (ViT) - a new class of deep learning architectures - have emerged in recent years as an alternative to CNNs for several computer vision applications. Here we tested variants of the ViT architecture for a range of desired neuroimaging downstream tasks based on difficulty, in this case for sex and Alzheimer's disease (AD) classification based on 3D brain MRI. In our experiments, two vision transformer architecture variants achieved an AUC of 0.987 for sex and 0.892 for AD classification, respectively. We independently evaluated our models on data from two benchmark AD datasets. We achieved a performance boost of 5% and 9-10% upon fine-tuning vision transformer models pre-trained on synthetic (generated by a latent diffusion model) and real MRI scans, respectively. Our main contributions include testing the effects of different ViT training strategies including pre-training, data augmentation and learning rate warm-ups followed by annealing, as pertaining to the neuroimaging domain. These techniques are essential for training ViT-like models for neuroimaging applications where training data is usually limited. We also analyzed the effect of the amount of training data utilized on the test-time performance of the ViT via data-model scaling curves.

Few-Shot Classification of Autism Spectrum Disorder using Site-Agnostic Meta-Learning and Brain MRI.

For machine learning applications in medical imaging, the availability of training data is often limited, which hampers the design of radiological classifiers for subtle conditions such as autism spectrum disorder (ASD). Transfer learning is one method to counter this problem of low training data regimes. Here we explore the use of meta-learning for very low data regimes in the context of having prior data from multiple sites - an approach we term site-agnostic meta-learning. Inspired by the effectiveness of meta-learning for optimizing a model across multiple tasks, here we propose a framework to adapt it to learn across multiple sites. We tested our meta-learning model for classifying ASD versus typically developing controls in 2,201 T1-weighted (T1-w) MRI scans collected from 38 imaging sites as part of Autism Brain Imaging Data Exchange (ABIDE) [age: 5.2-64.0 years]. The method was trained to find a good initialization state for our model that can quickly adapt to data from new unseen sites by fine-tuning on the limited data that is available. The proposed method achieved an ROC-AUC=0.857 on 370 scans from 7 unseen sites in ABIDE using a few-shot setting of 2-way 20-shot i.e., 20 training samples per site. Our results outperformed a transfer learning baseline by generalizing across a wider range of sites as well as other related prior work. We also tested our model in a zero-shot setting on an independent test site without any additional fine-tuning. Our experiments show the promise of the proposed site-agnostic meta-learning framework for challenging neuroimaging tasks involving multi-site heterogeneity with limited availability of training data.

Harnessing clinical annotations to improve deep learning performance in prostate segmentation.

PURPOSE: Developing large-scale datasets with research-quality annotations is challenging due to the high cost of refining clinically generated markup into high precision annotations. We evaluated the direct use of a large dataset with only clinically generated annotations in development of high-performance segmentation models for small research-quality challenge datasets. MATERIALS AND METHODS: We used a large retrospective dataset from our institution comprised of 1,620 clinically generated segmentations, and two challenge datasets (PROMISE12: 50 patients, ProstateX-2: 99 patients). We trained a 3D U-Net convolutional neural network (CNN) segmentation model using our entire dataset, and used that model as a template to train models on the challenge datasets. We also trained versions of the template model using ablated proportions of our dataset, and evaluated the relative benefit of those templates for the final models. Finally, we trained a version of the template model using an out-of-domain brain cancer dataset, and evaluated the relevant benefit of that template for the final models. We used five-fold cross-validation (CV) for all training and evaluation across our entire dataset. RESULTS: Our model achieves state-of-the-art performance on our large dataset (mean overall Dice 0.916, average Hausdorff distance 0.135 across CV folds). Using this model as a pre-trained template for refining on two external datasets significantly enhanced performance (30% and 49% enhancement in Dice scores respectively). Mean overall Dice and mean average Hausdorff distance were 0.912 and 0.15 for the ProstateX-2 dataset, and 0.852 and 0.581 for the PROMISE12 dataset. Using even small quantities of data to train the template enhanced performance, with significant improvements using 5% or more of the data. CONCLUSION: We trained a state-of-the-art model using unrefined clinical prostate annotations and found that its use as a template model significantly improved performance in other prostate segmentation tasks, even when trained with only 5% of the original dataset.

Optimizing Spatial Biopsy Sampling for the Detection of Prostate Cancer.

PURPOSE: The appropriate number of systematic biopsy cores to retrieve during magnetic resonance imaging (MRI)-targeted prostate biopsy is not well defined. We aimed to demonstrate a biopsy sampling approach that reduces required core count while maintaining diagnostic performance. MATERIALS AND METHODS: We collected data from a cohort of 971 men who underwent MRI-ultrasound fusion targeted biopsy for suspected prostate cancer. A regional targeted biopsy (RTB) was evaluated retrospectively; only cores within 2 cm of the margin of a radiologist-defined region of interest were considered part of the RTB. We compared detection rates for clinically significant prostate cancer (csPCa) and cancer upgrading rate on final whole mount pathology after prostatectomy between RTB, combined, MRI-targeted, and systematic biopsy. RESULTS: A total of 16,459 total cores from 971 men were included in the study data sets, of which 1,535 (9%) contained csPCa. The csPCa detection rates for systematic, MRI-targeted, combined, and RTB were 27.0% (262/971), 38.3% (372/971), 44.8% (435/971), and 44.0% (427/971), respectively. Combined biopsy detected significantly more csPCa than systematic and MRI-targeted biopsy (p <0.001 and p=0.004, respectively) but was similar to RTB (p=0.71), which used on average 3.8 (22%) fewer cores per patient. In 102 patients who underwent prostatectomy, there was no significant difference in upgrading rates between RTB and combined biopsy (p=0.84). CONCLUSIONS: A RTB approach can maintain state-of-the-art detection rates while requiring fewer retrieved cores. This result informs decision making about biopsy site selection and total retrieved core count.

Semi-automated PIRADS scoring via mpMRI analysis.

Purpose: Prostate cancer (PCa) is the most common solid organ cancer and second leading cause of death in men. Multiparametric magnetic resonance imaging (mpMRI) enables detection of the most aggressive, clinically significant PCa (csPCa) tumors that require further treatment. A suspicious region of interest (ROI) detected on mpMRI is now assigned a Prostate Imaging-Reporting and Data System (PIRADS) score to standardize interpretation of mpMRI for PCa detection. However, there is significant inter-reader variability among radiologists in PIRADS score assignment and a minimal input semi-automated artificial intelligence (AI) system is proposed to harmonize PIRADS scores with mpMRI data. Approach: The proposed deep learning model (the seed point model) uses a simulated single-click seed point as input to annotate the lesion on mpMRI. This approach is in contrast to typical medical AI-based approaches that require annotation of the complete lesion. The mpMRI data from 617 patients used in this study were prospectively collected at a major tertiary U.S. medical center. The model was trained and validated to classify whether an mpMRI image had a lesion with a PIRADS score greater than or equal to PIRADS 4. Results: The model yielded an average receiver-operator characteristic (ROC) area under the curve (ROC-AUC) of 0.704 over a 10-fold cross-validation, which is significantly higher than the previously published benchmark. Conclusions: The proposed model could aid in PIRADS scoring of mpMRI, providing second reads to promote quality as well as offering expertise in environments that lack a radiologist with training in prostate mpMRI interpretation. The model could help identify tumors with a higher PIRADS for better clinical management and treatment of PCa patients at an early stage.

Temporal Detection of Changes in the Vascularity and Concentration of Pigment Structures of a Skin Lesion.

Changes in morphology of a skin Iesion is indicative of melanoma, a deadly type of skin cancer. This paper proposes a temporal analysis approach to monitor the vascular appearance, the pigment structure, and growth of a skin Iesion. A set of digital images of a patients- skin Iesion acquired during follow-up imaging sessions serves as an input to our proposed system. The vascularity of the Iesion is modelled as the Kullback-Leibler (KL) divergence of the skin images- red channel information. The Iesion-s melanin pigment structures are quantified in terms of the textural energy and the ratio of its coverage to the total Iesion area. An optical flow field and related divergence field are implemented to indicate the direction of growth in a Iesion during follow-up image scans. An auto-regressive (AR) model predicts the change in the growth with time. Our results show the capability of the system proposed for real-time as well as off-line skin Iesion image analysis.