Search

BACKGROUND: Accurately detecting a variety of lung abnormalities from heterogenous chest X-ray (CXR) images and writing radiology reports is often difficult and time-consuming. OBJECTIVE: To access the utility of a novel artificial intelligence (AI) system (MOM-ClaSeg) in enhancing the accuracy and efficiency of radiologists in detecting heterogenous lung abnormalities through a multi-reader and multi-case (MRMC) observer performance study. METHODS: Over 36,000 CXR images were retrospectively collected from 12 hospitals over 4 months and used as the experiment group and the control group. In the control group, a double reading method is used in which two radiologists interpret CXR to generate a final report, while in the experiment group, one radiologist generates the final reports based on AI-generated reports. RESULTS: Compared with double reading, the diagnostic accuracy and sensitivity of single reading with AI increases significantly by 1.49% and 10.95%, respectively (P < 0.001), while the difference in specificity is small (0.22%) and without statistical significance (P = 0.255). Additionally, the average image reading and diagnostic time in the experimental group is reduced by 54.70% (P < 0.001). CONCLUSION: This MRMC study demonstrates that MOM-ClaSeg can potentially serve as the first reader to generate the initial diagnostic reports, with a radiologist only reviewing and making minor modifications (if needed) to arrive at the final decision. It also shows that single reading with AI can achieve a higher diagnostic accuracy and efficiency than double reading. [ABSTRACT FROM AUTHOR]

Radar sensors, leveraging the Doppler effect, enable the nonintrusive capture of kinetic and physiological motions while preserving privacy. Deep learning (DL) facilitates radar sensing for healthcare applications such as gait recognition and vital-sign measurement. However, band-dependent patterns, indicating variations in patterns and power scales associated with frequencies in time–frequency representation (TFR), challenge radar sensing applications using DL. Frequency-dependent characteristics and features with lower power scales may be overlooked during representation learning. This paper proposes an Enhanced Band-Dependent Learning framework (E-BDL) comprising an adaptive sub-band filtering module, a representation learning module, and a sub-view contrastive module to fully detect band-dependent features in sub-frequency bands and leverage them for classification. Experimental validation is conducted on two radar datasets, including gait abnormality recognition for Alzheimer's disease (AD) and AD-related dementia (ADRD) risk evaluation and vital-sign monitoring for hemodynamics scenario classification. For hemodynamics scenario classification, E-BDL-ResNet achieves competitive performance in overall accuracy and class-wise evaluations compared to recent methods. For ADRD risk evaluation, the results demonstrate E-BDL-ResNet's superior performance across all candidate models, highlighting its potential as a clinical tool. E-BDL effectively detects salient sub-bands in TFRs, enhancing representation learning and improving the performance and interpretability of DL-based models. [ABSTRACT FROM AUTHOR]

Machine learning has shown great promise for integrating multi-modality neuroimaging datasets to predict the risk of progression/conversion to Alzheimer's Disease (AD) for individuals with Mild Cognitive Impairment (MCI). Most existing work aims to classify MCI patients into converters versus non-converters using a pre-defined timeframe. The limitation is a lack of granularity in differentiating MCI patients who convert at different paces. Progression pace prediction has important clinical values, which allow from more personalized interventional strategies, better preparation of patients and their caregivers, and facilitation of patient selection in clinical trials. We proposed a novel ADPacer model which formulated the pace prediction into an ordinal learning problem with a unique capability of leveraging training samples with label ambiguity to augment the training set. This capability differentiates ADPacer from existing ordinal learning algorithms. We applied ADPacer to MCI patient cohorts from the Alzheimer's Disease Neuroimaging Initiative (ADNI) and the Australian Imaging, Biomarker & Lifestyle Flagship Study of Aging (AIBL), and demonstrated the superior performance of ADPacer compared to existing ordinal learning algorithms. We also integrated the SHapley Additive exPlanations (SHAP) method with ADPacer to assess the contributions from different modalities to the model prediction. The findings are consistent with the AD literature. [ABSTRACT FROM AUTHOR]

Alzheimer's disease (AD) is a devastating neurodegenerative disorder. Early prediction of the risk of converting to AD for individuals at pre-dementia stages such as Mild Cognitive Impairment (MCI) is important. This could provide an opportunity for early intervention to slow down disease progression before significant irreversible neurodegeneration occurs. Neuroimaging datasets of different modalities such as MRI and PET have shown great promise. However, different data modalities are associated with varying acquisition costs/levels of accessibility to patients. We propose a machine learning (ML) framework, namely Uncertainty-driven Modality Selection (UMoS), that allows for sequentially adding data modalities for each patient on an as-needed basis, while at the same time achieving high prediction accuracy as if all the modalities were used. UMoS provides a tool to assist clinicians in deciding what data modalities/diagnostic exams each patient needs. We apply UMoS to a real-world dataset from the Alzheimer's Disease Neuroimaging Initiative (ADNI) based on demographic/clinical data, MRI, and PET. UMoS shows high accuracy for predicting MCI conversion to AD, which has no significant difference from models based on the simultaneous use of all the modalities for each patient. The benefit of UMoS is significant data efficiency accomplished by saving a large percentage of patients from needing to acquire more costly/less accessible data modalities, thus lessening the burden on patients and the healthcare system. [ABSTRACT FROM AUTHOR]

Early diagnosis of Alzheimer's disease (AD) is an important task that facilitates the development of treatment and prevention strategies, and may potentially improve patient outcomes. Neuroimaging has shown great promise, including the amyloid-PET, which measures the accumulation of amyloid plaques in the brain—a hallmark of AD. It is desirable to train end-to-end deep learning models to predict the progression of AD for individuals at early stages based on 3D amyloid-PET. However, commonly used models are trained in a fully supervised learning manner, and they are inevitably biased toward the given label information. To this end, we propose a selfsupervised contrastive learning method to accurately predict the conversion to AD for individuals with mild cognitive impairment (MCI) with 3D amyloid-PET. The proposed method, SMoCo, uses both labeled and unlabeled data to capture general semantic representations underlying the images. As the downstream task is given as classification of converters vs. non-converters, unlike the general self-supervised learning problem that aims to generate task-agnostic representations, SMoCo additionally utilizes the label information in the pre-training. To demonstrate the performance of our method, we conducted experiments on the Alzheimer's Disease Neuroimaging Initiative (ADNI) dataset. The results confirmed that the proposed method is capable of providing appropriate data representations, resulting in accurate classification. SMoCo showed the best classification performance over the existing methods, with AUROC = 85.17%, accuracy = 81.09%, sensitivity = 77.39%, and specificity = 82.17%. While SSL has demonstrated great success in other application domains of computer vision, this study provided the initial investigation of using a proposed self-supervised contrastive learning model, SMoCo, to effectively predict MCI conversion to AD based on 3D amyloid-PET. [ABSTRACT FROM AUTHOR]

Multimodality datasets are becoming increasingly common in various domains to provide complementary information for predictive analytics. One significant challenge in fusing multimodality data is that the multiple modalities are not universally available for all samples due to cost and accessibility constraints. This situation results in a unique data structure called an Incomplete Multimodality Dataset. We propose a novel Incomplete-Multimodality Transfer Learning (IMTL) model that builds a predictive model for each sub-cohort of samples with the same missing modality pattern, and meanwhile couples the model estimation processes for different sub-cohorts to allow for transfer learning. We develop an Expectation-Maximization (EM) algorithm to estimate the parameters of IMTL and further extend it to a collaborative learning paradigm that is specifically valuable for patient privacy preservation in health care applications. We prove two advantageous properties of IMTL: the ability for out-of-sample prediction and a theoretical guarantee for a larger Fisher information compared with models without transfer learning. IMTL is applied to diagnosis and prognosis of Alzheimer's disease at an early stage called Mild Cognitive Impairment using incomplete multimodality imaging data. IMTL achieves higher accuracy than competing methods without transfer learning. [ABSTRACT FROM AUTHOR]