Artificial intelligence (AI) (decision trees, regression algorithms, support vector machines, Bayesian methods, neural networks, etc.) has been utilized for decades to address a variety of medical imaging problems, such as image segmentation¹ (i.e., finding the borders of a target object), registration² (i.e., visually aligning anatomical parts in single- or multimodality images), detection (i.e., detecting formations/structures), and classification³ (i.e., grouping of medical information in subgroups). It can also facilitate information feeds in radiology workflows^4,5 (e.g., natural language processing in dictation systems). “Machine learning” (ML) is an application of AI in which the computer (machine) is given data access, and models are used for extracting relevant information from the data. The recent usage of neural networks, which is a well-established ML approach, has gained significant momentum with generalization of the techniques to deeper network architectures, referred to as deep neural networks (DNNs); the complete concept is known as “deep learning” (DL).⁶ Research on deeper architectures shows that the accuracy of the deployed models depends heavily on the amount of relevant information; therefore, access to both past data and the ongoing inflow of new information is critical. Accordingly, DL-based solutions are commonly built on vast amounts of data.⁷

Brain Metastases (BM) complicate 20-40% of cancer cases. BM lesions can present as punctate (1 mm) foci, requiring high-precision Magnetic Resonance Imaging (MRI) in order to prevent inadequate or delayed BM treatment. However, BM lesion detection remains challenging partly due to their structural similarities to normal structures (e.g., vasculature). We propose a BM-detection framework using a single-sequence gadolinium-enhanced T1-weighted 3D MRI dataset. The framework focuses on the detection of smaller (< 15 mm) BM lesions and consists of: (1) candidate-selection stage, using Laplacian of Gaussian approach for highlighting parts of an MRI volume holding higher BM occurrence probabilities, and (2) detection stage that iteratively processes cropped region-of-interest volumes centered by candidates using a custom-built 3D convolutional neural network (“CropNet”). Data is augmented extensively during training via a pipeline consisting of random gamma correction and elastic deformation stages; the framework thereby maintains its invariance for a plausible range of BM shape and intensity representations. This approach is tested using five-fold cross-validation on 217 datasets from 158 patients, with training and testing groups randomized per patient to eliminate learning bias. The BM database included lesions with a mean diameter of ~5.4 mm and a mean volume of ~160 mm3. For 90% BM-detection sensitivity, the framework produced on average 9.12 false-positive BM detections per patient (standard deviation of 3.49); for 85% sensitivity, the average number of false-positives declined to 5.85. Comparative analysis showed that the framework produces comparable BM-detection accuracy with the state-of-art approaches validated for significantly larger lesions.

<!-- *** Custom HTML *** --> Purpose: Sharing medical images between institutions, or even inside the same institution, is restricted by various laws and regulations; research projects requiring large datasets may suffer as a result. These limitations might be addressed by an abundant supply of synthetic data that (1) are representative (i.e., the synthetic data could produce comparable research results as the original data) and (2) do not closely resemble the original images (i.e., patient privacy is protected). We introduce a framework that generates data with these requirements leveraging generative adversarial network (GAN) ensembles in a controlled fashion. Approach: To this end, an adaptive ensemble scaling strategy with the objective of representativeness is defined. A sampled Fréchet distance-based constraint was then created to eliminate poorly converged candidates. Finally, a mutual information-based validation metric was embedded into the framework to confirm there are visual differences between the original and the generated synthetic images. Results: The applicability of the solution is demonstrated with a case study for generating three-dimensional brain metastasis (BM) from T1-weighted contrast-enhanced MRI studies. A previously published BM detection system was reported to produce 9.12 false-positives at 90% detection sensitivity based on the original data. By using the synthetic data generated with the proposed framework, the system produced 9.53 false-positives at the same sensitivity level. Conclusions: Achieving comparable algorithm performance relying solely on synthetic data unveils a significant potential to eliminate/reduce patient privacy concerns when sharing data in medical imaging.

The detection of brain metastases (BM) in their early stages could have a positive impact on the outcome of cancer patients. The authors previously developed a framework for detecting small BM (with diameters of 10%) were also observed in reduced labeled data scenarios (using 50% and 75% of labeled data). The results suggest that the introduced strategy could be utilized in existing medical detection applications with access to unlabeled datasets to elevate their performances.

The emergence of the COVID-19 pandemic over a relatively brief interval illustrates the need for rapid data-driven approaches to facilitate clinical decision making. We examined a machine learning process to predict inpatient mortality among COVID-19 patients using clinical and chest radiographic data. Modeling was performed with a de-identified dataset of encounters prior to widespread vaccine availability. Non-imaging predictors included demographics, pre-admission clinical history, and past medical history variables. Imaging features were extracted from chest radiographs by applying a deep convolutional neural network with transfer learning. A multi-layer perceptron combining 64 deep learning features from chest radiographs with 98 patient clinical features was trained to predict mortality. The Local Interpretable Model-Agnostic Explanations (LIME) method was used to explain model predictions. Non-imaging data alone predicted mortality with an ROC-AUC of 0.87 ± 0.03 (mean ± SD), while the addition of imaging data improved prediction slightly (ROC-AUC: 0.91 ± 0.02). The application of LIME to the combined imaging and clinical model found HbA1c values to contribute the most to model prediction (17.1 ± 1.7%), while imaging contributed 8.8 ± 2.8%. Age, gender, and BMI contributed 8.7%, 8.2%, and 7.1%, respectively. Our findings demonstrate a viable explainable AI approach to quantify the contributions of imaging and clinical data to COVID mortality predictions.