Keywords: Diagnosis/Prediction, Aging, Ex-Vivo Applications, Brain, Microbleeds

Motivation: Accurate and efficient detection of cerebral microbleeds (CMBs) on postmortem MRI is necessary for MR-pathology studies on the relationship between CMBs and cerebral small vessel disease (SVD). 

Goal(s): The development and improvement of an automated detection framework for identifying cerebral microbleeds (CMBs) on MRI scans of community-based older adults.

Approach: Fuzzy segmentation, a novel self-supervised auxiliary task based on CMB data synthesis, is proposed for pre-training a CMB detection model alongside other state-of-the-art SSL methods.

Results: Self-supervised pre-training with fuzzy segmentation and rotation prediction led to an 11% increase in average precision for automated CMB detection on postmortem MRI. 

Impact: This study demonstrates a new state-of-the-art for postmortem CMB detection performance using self-supervised learning. Automated CMB detection on postmortem MRI will enable future MR-pathology studies into the links between CMBs and neuropathology observed at autopsy such as cerebral amyloid angiopathy.

Keywords: Diagnosis/Prediction, Machine Learning/Artificial Intelligence, microbleed, ARIA, SWI

Motivation: Cerebral microbleeds (CMBs) are small brain hemorrhages detectable with MRI associated with conditions like cerebral amyloid angiopathy. As their detection can be difficult, automated methods are needed for quick and precise detection and localization of CMBs.

Goal(s): To propose an algorithm to detect CMBs.

Approach: A neural network was trained on SWI/T2* images, with artificial bleeds generated and added during training. The model’s performance was tested on an independent test set with actual CMBs.

Results: Despite the absence of real CMBs in the training data, the simulated bleeds provided sufficient information to train a model with good performance in the independent test set.

Impact: We propose an algorithm that can help with the tedious radiological task of detecting cerebral microbleeds in the brain. We further demonstrate that a model trained solely on simulated bleeds can effectively detect actual microbleeds in real MRI data.

Keywords: Diagnosis/Prediction, Multiple Sclerosis, Deep Learning, Segmentation, Spinal Cord

Motivation: Longitudinal analysis of spinal cord multiple sclerosis (MS) lesions is clinically relevant for the early diagnosis and monitoring of MS progression.

Goal(s): Develop a deep learning tool for the automatic segmentation of MS spinal cord lesions on PSIR and STIR images from multiple sites.

Approach: A nnUNet model was trained and tested on the baseline data and applied to follow-up scans to create lesion distribution maps.

Results: We demonstrated the utility of the model to map the spatio-temporal distribution of MS lesions across MS phenotypes. The model is packaged into an open-source software.

Impact: Automatic segmentation of spinal cord lesions in large cohorts helps to identify signatures of MS phenotypes for ultimately improving prognosis and optimizing treatment for people with MS.

Keywords: AI/ML Image Reconstruction, Ischemia

Motivation: The diagnostic performance of portable low-field-strength MRI (LF-MRI) is constrained by low spatial-resolution and signal-to-noise ratio.

Goal(s): To evaluate the performance in detecting and quantifying ischemic lesions among SynthMRI, LF-MRI and real high-field-strength MRI (HF-MRI).

Approach: We created a deep learning-based model to generate the synthetic super-resolution (3T) MRI (SynthMRI) based on LF-MRI (0.23T). We evaluated the performance in detecting and quantifying ischemic lesions among SynthMRI, LF-MRI and HF-MRI.

Results: SynthMRI demonstrated high sensitivity in detecting the number and locations of ischemic lesions. Moreover, SynthMRI exhibited strong correlations with HF-MRI in the quantitative assessment of ischemic lesions, and significantly higher than portable LF-MRI.

Impact: Synthetic super-resolution MRI images overcome the limitations of low spatial resolution and signal-to-noise ratio in portable low-field-strength MRI. It has the potential to replace high-field-strength MRI images in the neuroimaging of AIS, enabling portable low-field-strength MRI examinations with comparable performance.

Keywords: Diagnosis/Prediction, Brain

Motivation: Deep learning-based segmentation methods have shown promising results; however, they require a large number of segmentation labels for training, which is very costly to obtain, especially for 3D labels.

Goal(s): Our goal is to achieve promising 3D segmentation results with few labels by exploiting the ability to capture semantic information from 2D diffusion models trained without labels.

Approach: We train simple pixel classifiers using features extracted from 2D diffusion models that have been trained with slices from three orthogonal orientations.

Results: In our experiments on the Human Connectome Project database, our proposed method outperformed conventional segmentation methods in a few labeled scenarios.

Impact: Our proposed method for segmenting subcortical brain structures can be readily applied to pre-trained diffusion models with only a few labeled data, while also generating paired segmentation labels for the images produced by diffusion models.

Keywords: AI Diffusion Models, Segmentation, Liver Vessel Segmentation

Motivation: To improve liver vessel segmentation on MRI under annotation constraints.

Goal(s): Apply an advanced unpaired image translation technique, SynDiff, to create synthetic MR images from CT data.

Approach: By incorporating vessel masks in the translation process, the optimized SynDiff models generated synthetic images that facilitated more effective pretraining of segmentation models.

Results: Validated across multiple pretraining settings, the refined SynDiff approach surpassed the standard nnU-Net and other pretraining-based methods, substantially improving liver vessel segmentation performance.

Impact: This study remarkably advances liver vessel segmentation on MRI, demonstrating that synthetic data can effectively augment limited datasets, leading to improved model performance. It has great potential for broader applications in medical image analysis.

Keywords: Analysis/Processing, Alzheimer's Disease, multimodal, language-text

Motivation: The diagnosis of Alzheimer's disease (AD) considers not only clinical symptoms but also various data sources, including MR imaging.

Goal(s): In this study, we used a multimodal approach to integrate both language and vision information to improve the performance of clinical dementia rating classification network.

Approach: We used contrastive pre-training with language and vision data pairs; then trained a classifier, freezing the pre-trained network during classifier training.

Results: The results show that the integrated model achieved the highest accuracy. In addition, the contrastive learning process improved the performance of the vision encoder with guidance of abundant linguistic information.

Impact: With multimodal training, we successfully integrated both vision and language information and yielded the best results with integrated model. Also, multimodal training enhanced vision encoder's performance. When limited language information was provided, the complementary information from visual information was greater.

Keywords: AI Diffusion Models, Machine Learning/Artificial Intelligence, Oncology, Cancer, DMG, Diffuse Midline Glioma

Motivation: Pediatric diffuse midline gliomas are associated with a poor prognosis, leaving radiotherapy as standard of palliative care. Personalized radiation regimes could maximize the benefit for the patient, and consequently improve clinical outcomes.

Goal(s): This study explores a state-of-the-art computer vision method to predict the anatomical growth of tumors which could inform tailored radiotherapy treatments.

Approach: A denoising diffusion implicit model is employed to generate realistic, high-quality magnetic resonance imaging scans of enlarged tumor sizes starting from a baseline image.

Results: Our proof-of-concept study demonstrates promising results on an external longitudinal pediatric dataset, highlighting the method’s potential to realistically predict visual tumor growth. 

Impact:  We demonstrate realistic predictions of anatomical (pediatric) brain tumor growth using a generative denoising diffusion implicit model. This enables personalized predictions of tumor growth trajectories to guide localized therapies such as geometric dose shaping for radiotherapy delivery.

Keywords: Analysis/Processing, Aging, Biological Age

Motivation: MRI is a valuable tool for providing health-related information, including visualizing age-associated changes. Aging is influenced by chronic diseases, and assessing the true organ-specific biological age is essential for accurate diagnosis.

Goal(s): Development of an organ-specific age estimation for investigation in large imaging cohort with associated non-imaging information.

Approach: While prior studies focused on age estimation from non-imaging data or single organs, we propose a multi-organ age estimation framework, operating on brain, cardiac, and abdominal MRIs, and OCT scans.

Results: Our results prove the feasibility of imaging-based organ age estimation and initiate further investigations to identify risk factors for accelerated aging.

Impact: Reliable imaging-based estimation of biological age in multiple organ systems facilitates research efforts to identify risk factors of accelerated aging, advancing the goal of age-related phenotyping.

Keywords: Analysis/Processing, Machine Learning/Artificial Intelligence

Motivation: Although deep learning frameworks have been widely used in all aspects of the MR imaging pipeline, the effect of learning tissue-specific information from MR images in improving model performance needs to be understood.

Goal(s): We demonstrate the utility of a self-supervised contrastive learning framework that uses multi-contrast information to improve synthesis of T1w and T2w images.

Approach: A deep learning model is pretrained to learn T1 and T2 information from a set of multi-parametric MR images.

Results: A contrast synthesis framework was developed using few examples of contrast mapping. Embedding relevant contrast information during pretraining synthesized images with improved MSE, SSIM, and PSNR. 

Impact: Multi-contrast information can be leveraged by self-supervised deep learning models to understand underlying tissue characteristics and synthesize new MR contrast-weighted images. This demonstrates the wider applicability of embedding tissue-specific information in improving different aspects of the MR imaging pipeline.