ABSTRACT
BACKGROUND: Deep learning has demonstrated significant advancements across various domains. However, its implementation in specialized areas, such as medical settings, remains approached with caution. In these high-stake environments, understanding the model's decision-making process is critical. This study assesses the performance of different pretrained Bidirectional Encoder Representations from Transformers (BERT) models and delves into understanding its decision-making within the context of medical image protocol assignment. METHODS: Four different pre-trained BERT models (BERT, BioBERT, ClinicalBERT, RoBERTa) were fine-tuned for the medical image protocol classification task. Word importance was measured by attributing the classification output to every word using a gradient-based method. Subsequently, a trained radiologist reviewed the resulting word importance scores to assess the model's decision-making process relative to human reasoning. RESULTS: The BERT model came close to human performance on our test set. The BERT model successfully identified relevant words indicative of the target protocol. Analysis of important words in misclassifications revealed potential systematic errors in the model. CONCLUSIONS: The BERT model shows promise in medical image protocol assignment by reaching near human level performance and identifying key words effectively. The detection of systematic errors paves the way for further refinements to enhance its safety and utility in clinical settings.
Subject(s)
Natural Language Processing , Problem Solving , HumansABSTRACT
Perspiration-based wearable biosensors facilitate continuous monitoring of individuals' health states with real-time and molecular-level insight. The inherent inaccessibility of sweat in sedentary individuals in large volume (≥10 µL) for on-demand and in situ analysis has limited our ability to capitalize on this noninvasive and rich source of information. A wearable and miniaturized iontophoresis interface is an excellent solution to overcome this barrier. The iontophoresis process involves delivery of stimulating agonists to the sweat glands with the aid of an electrical current. The challenge remains in devising an iontophoresis interface that can extract sufficient amount of sweat for robust sensing, without electrode corrosion and burning/causing discomfort in subjects. Here, we overcame this challenge through realizing an electrochemically enhanced iontophoresis interface, integrated in a wearable sweat analysis platform. This interface can be programmed to induce sweat with various secretion profiles for real-time analysis, a capability which can be exploited to advance our knowledge of the sweat gland physiology and the secretion process. To demonstrate the clinical value of our platform, human subject studies were performed in the context of the cystic fibrosis diagnosis and preliminary investigation of the blood/sweat glucose correlation. With our platform, we detected the elevated sweat electrolyte content of cystic fibrosis patients compared with that of healthy control subjects. Furthermore, our results indicate that oral glucose consumption in the fasting state is followed by increased glucose levels in both sweat and blood. Our solution opens the possibility for a broad range of noninvasive diagnostic and general population health monitoring applications.
Subject(s)
Cystic Fibrosis/metabolism , Glucose/metabolism , Sweat/metabolism , Wearable Electronic Devices , Humans , Iontophoresis/instrumentation , Iontophoresis/methods , Monitoring, Physiologic/instrumentation , Monitoring, Physiologic/methodsABSTRACT
Dynamic susceptibility contrast magnetic resonance perfusion (DSC-MRP) is a non-invasive imaging technique for hemodynamic measurements. Various perfusion parameters, such as cerebral blood volume (CBV) and cerebral blood flow (CBF), can be derived from DSC-MRP, hence this non-invasive imaging protocol is widely used clinically for the diagnosis and assessment of intracranial pathologies. Currently, most institutions use commercially available software to compute the perfusion parametric maps. However, these conventional methods often have limitations, such as being time-consuming and sensitive to user input, which can lead to inconsistent results; this highlights the need for a more robust and efficient approach like deep learning. Using the relative cerebral blood volume (rCBV) and relative cerebral blood flow (rCBF) perfusion maps generated by FDA-approved software, we trained a multistage deep learning model. The model, featuring a combination of a 1D convolutional neural network (CNN) and a 2D U-Net encoder-decoder network, processes each 4D MRP dataset by integrating temporal and spatial features of the brain for voxel-wise perfusion parameters prediction. An auxiliary model, with similar architecture, but trained with truncated datasets that had fewer time-points, was designed to explore the contribution of temporal features. Both qualitatively and quantitatively evaluated, deep learning-generated rCBV and rCBF maps showcased effective integration of temporal and spatial data, producing comprehensive predictions for the entire brain volume. Our deep learning model provides a robust and efficient approach for calculating perfusion parameters, demonstrating comparable performance to FDA-approved commercial software, and potentially mitigating the challenges inherent to traditional techniques.
Subject(s)
Cerebral Blood Volume , Cerebrovascular Circulation , Deep Learning , Humans , Cerebrovascular Circulation/physiology , Cerebral Blood Volume/physiology , Magnetic Resonance Imaging/methods , Male , Brain/blood supply , Brain/diagnostic imaging , Female , AdultABSTRACT
Diagnosis of endoleak following endovascular aortic repair (EVAR) relies on manual review of multi-slice CT angiography (CTA) by physicians which is a tedious and time-consuming process that is susceptible to error. We evaluate the use of a deep neural network for the detection of endoleak on CTA for post-EVAR patients using a novel data efficient training approach. 50 CTAs and 20 CTAs with and without endoleak respectively were identified based on gold standard interpretation by a cardiovascular subspecialty radiologist. The Endoleak Augmentor, a custom designed augmentation method, provided robust training for the machine learning (ML) model. Predicted segmentation maps underwent post-processing to determine the presence of endoleak. The model was tested against 3 blinded general radiologists and 1 blinded subspecialist using a held-out subset (10 positive endoleak CTAs, 10 control CTAs). Model accuracy, precision and recall for endoleak diagnosis were 95%, 90% and 100% relative to reference subspecialist interpretation (AUC = 0.99). Accuracy, precision and recall was 70/70/70% for generalist1, 50/50/90% for generalist2, and 90/83/100% for generalist3. The blinded subspecialist had concordant interpretations for all test cases compared with the reference. In conclusion, our ML-based approach has similar performance for endoleak diagnosis relative to subspecialists and superior performance compared with generalists.