Population Based Image Imputation

Simulated MRI Artifacts: Testing Machine Learning Failure Modes

Objective. Seven types of MRI artifacts, including acquisition and preprocessing errors, were simulated to test a machine learning brain tumor segmentation model for potential failure modes. Introduction. Real-world medical deployments of machine learning algorithms are less common than the number of medical research papers using machine learning. Part of the gap between the performance of models in research and deployment comes from a lack of hard test cases in the data used to train a model. Methods. These failure modes were simulated for a pretrained brain tumor segmentation model that utilizes standard MRI and used to evaluate the performance of the model under duress. These simulated MRI artifacts consisted of motion, susceptibility induced signal loss, aliasing, field inhomogeneity, sequence mislabeling, sequence misalignment, and skull stripping failures. Results. The artifact with the largest effect was the simplest, sequence mislabeling, though motion, field inhomogeneity, and sequence misalignment also caused significant performance decreases. The model was most susceptible to artifacts affecting the FLAIR (fluid attenuation inversion recovery) sequence. Conclusion. Overall, these simulated artifacts could be used to test other brain MRI models, but this approach could be used across medical imaging applications.

SMORE: A Self-Supervised Anti-Aliasing and Super-Resolution Algorithm for MRI Using Deep Learning

High resolution magnetic resonance (MR) images are desired in many clinical and research applications. Acquiring such images with high signal-to-noise (SNR), however, can require a long scan duration, which is difficult for patient comfort, is more costly, and makes the images susceptible to motion artifacts. A very common practical compromise for both 2D and 3D MR imaging protocols is to acquire volumetric MR images with high in-plane resolution, but lower through-plane resolution. In addition to having poor resolution in one orientation, 2D MRI acquisitions will also have aliasing artifacts, which further degrade the appearance of these images. This paper presents an approach SMORE1 based on convolutional neural networks (CNNs) that restores image quality by improving resolution and reducing aliasing in MR images.2 This approach is self-supervised, which requires no external training data because the high-resolution and low-resolution data that are present in the image itself are used for training. For 3D MRI, the method consists of only one self-supervised super-resolution (SSR) deep CNN that is trained from the volumetric image data. For 2D MRI, there is a self-supervised anti-aliasing (SAA) deep CNN that precedes the SSR CNN, also trained from the volumetric image data. Both methods were evaluated on a broad collection of MR data, including filtered and downsampled images so that quantitative metrics could be computed and compared, and actual acquired low resolution images for which visual and sharpness measures could be computed and compared. The super-resolution method is shown to be visually and quantitatively superior to previously reported methods.

Medical Informatics Platform (MIP): A Pilot Study Across Clinical Italian Cohorts

Introduction: With the shift of research focus to personalized medicine in Alzheimer's Dementia (AD), there is an urgent need for tools that are capable of quantifying a patient's risk using diagnostic biomarkers. The Medical Informatics Platform (MIP) is a distributed e-infrastructure federating large amounts of data coupled with machine-learning (ML) algorithms and statistical models to define the biological signature of the disease. The present study assessed (i) the accuracy of two ML algorithms, i.e., supervised Gradient Boosting (GB) and semi-unsupervised 3C strategy (Categorize, Cluster, Classify-CCC) implemented in the MIP and (ii) their contribution over the standard diagnostic workup. Methods: We examined individuals coming from the MIP installed across 3 Italian memory clinics, including subjects with Normal Cognition (CN, n = 432), Mild Cognitive Impairment (MCI, n = 456), and AD (n = 451). The GB classifier was applied to best discriminate the three diagnostic classes in 1,339 subjects, and the CCC strategy was used to refine the classical disease categories. Four dementia experts provided their diagnostic confidence (DC) of MCI conversion on an independent cohort of 38 patients. DC was based on clinical, neuropsychological, CSF, and structural MRI information and again with addition of the outcome from the MIP tools. Results: The GB algorithm provided a classification accuracy of 85% in a nested 10-fold cross-validation for CN vs. MCI vs. AD discrimination. Accuracy increased to 95% in the holdout validation, with the omission of each Italian clinical cohort out in turn. CCC identified five homogeneous clusters of subjects and 36 biomarkers that represented the disease fingerprint. In the DC assessment, CCC defined six clusters in the MCI population used to train the algorithm and 29 biomarkers to improve patients staging. GB and CCC showed a significant impact, evaluated as +5.99% of increment on physicians' DC. The influence of MIP on DC was rated from "slight" to "significant" in 80% of the cases. Discussion: GB provided fair results in classification of CN, MCI, and AD. CCC identified homogeneous and promising classes of subjects via its semi-unsupervised approach. We measured the effect of the MIP on the physician's DC. Our results pave the way for the establishment of a new paradigm for ML discrimination of patients who will or will not convert to AD, a clinical priority for neurology.

Medical Image Imputation from Image Collections

We present an algorithm for creating high resolution anatomically plausible images consistent with acquired clinical brain MRI scans with large inter-slice spacing. Although large data sets of clinical images contain a wealth of information, time constraints during acquisition result in sparse scans that fail to capture much of the anatomy. These characteristics often render computational analysis impractical as many image analysis algorithms tend to fail when applied to such images. Highly specialized algorithms that explicitly handle sparse slice spacing do not generalize well across problem domains. In contrast, we aim to enable application of existing algorithms that were originally developed for high resolution research scans to significantly undersampled scans. We introduce a generative model that captures fine-scale anatomical structure across subjects in clinical image collections and derive an algorithm for filling in the missing data in scans with large inter-slice spacing. Our experimental results demonstrate that the resulting method outperforms state-of-the-art upsampling super-resolution techniques, and promises to facilitate subsequent analysis not previously possible with scans of this quality. Our implementation is freely available at https://github.com/adalca/papago.

Machine Learning in Radiology: Applications Beyond Image Interpretation

Much attention has been given to machine learning and its perceived impact in radiology, particularly in light of recent success with image classification in international competitions. However, machine learning is likely to impact radiology outside of image interpretation long before a fully functional "machine radiologist" is implemented in practice. Here, we describe an overview of machine learning, its application to radiology and other domains, and many cases of use that do not involve image interpretation. We hope that better understanding of these potential applications will help radiology practices prepare for the future and realize performance improvement and efficiency gains.