A recurrent neural network classifier for Doppler ultrasound blood flow signals

Evaluation of Deep Learning Model Architectures for Point-of-Care Ultrasound Diagnostics

Point-of-care ultrasound imaging is a critical tool for patient triage during trauma for diagnosing injuries and prioritizing limited medical evacuation resources. Specifically, an eFAST exam evaluates if there are free fluids in the chest or abdomen but this is only possible if ultrasound scans can be accurately interpreted, a challenge in the pre-hospital setting. In this effort, we evaluated the use of artificial intelligent eFAST image interpretation models. Widely used deep learning model architectures were evaluated as well as Bayesian models optimized for six different diagnostic models: pneumothorax (i) B- or (ii) M-mode, hemothorax (iii) B- or (iv) M-mode, (v) pelvic or bladder abdominal hemorrhage and (vi) right upper quadrant abdominal hemorrhage. Models were trained using images captured in 27 swine. Using a leave-one-subject-out training approach, the MobileNetV2 and DarkNet53 models surpassed 85% accuracy for each M-mode scan site. The different B-mode models performed worse with accuracies between 68% and 74% except for the pelvic hemorrhage model, which only reached 62% accuracy for all model architectures. These results highlight which eFAST scan sites can be easily automated with image interpretation models, while other scan sites, such as the bladder hemorrhage model, will require more robust model development or data augmentation to improve performance. With these additional improvements, the skill threshold for ultrasound-based triage can be reduced, thus expanding its utility in the pre-hospital setting.

Wavelet Scattering Transform based Doppler signal classification

Normal blood supply to the human brain may be marred by the presence of a clot inside the blood vessels. This clot structure called emboli inhibits normal blood flow to the brain. It is considered as one of the main sources of stroke. Presence of emboli in human's can be determined by the analysis of transcranial Doppler signal. Different signal processing and machine learning algorithms have been used for classifying the detected signal as an emboli, Doppler speckle, and an artifact. In this paper, we sought to make use of the wavelet transform based algorithm called Wavelet Scattering Transform, which is translation invariant and stable to deformations for classifying different Doppler signals. With its architectural resemblance to Convolutional Neural Network, Wavelet Scattering Transform works well on small datasets and subsequently was trained on a dataset consisting of 300 Doppler signals. To check the effectiveness of extracted Scattering transform based features for Doppler signal classification, learning algorithms that included multi-class Support vector machine, k-nearest neighbor and Naive Bayes algorithms were trained. Comparative analysis was done with respect to the handcrafted Continuous wavelet transform features extracted from samples and Wavelet scattering with Support vector machine achieved an accuracy of 98.89%. Also, with set of extracted scattering coefficients, Gaussian process regression was performed and a regression model was trained on three different sets of scattering coefficients with zero order scattering coefficients providing least prediction loss of 34.95%.

Modeling strategies to analyse longitudinal biomarker data: An illustration on predicting immunotherapy non-response in non-small cell lung cancer

Serum tumor markers acquired through a blood draw are known to reflect tumor activity. Their non-invasive nature allows for more frequent testing compared to traditional imaging methods used for response evaluations. Our study aims to compare nine prediction methods to accurately, and with a low false positive rate, predict progressive disease despite treatment (i.e. non-response) using longitudinal tumor biomarker data. Bi-weekly measurements of CYFRA, CA-125, CEA, NSE, and SCC were available from a cohort of 412 advanced stage non-small cell lung cancer (NSCLC) patients treated up to two years with immune checkpoint inhibitors. Serum tumor marker measurements from the first six weeks after treatment initiation were used to predict treatment response at 6 months. Nine models with varying complexity were evaluated in this study, showing how longitudinal biomarker data can be used to predict non-response to immunotherapy in NSCLC patients.

Biomedical system based on the Discrete Hidden Markov Model using the Rocchio-Genetic approach for the classification of internal carotid artery Doppler signals

When the maximum likelihood approach (ML) is used during the calculation of the Discrete Hidden Markov Model (DHMM) parameters, DHMM parameters of the each class are only calculated using the training samples (positive training samples) of the same class. The training samples (negative training samples) not belonging to that class are not used in the calculation of DHMM model parameters. With the aim of supplying that deficiency, by involving the training samples of all classes in calculating processes, a Rocchio algorithm based approach is suggested. During the calculation period, in order to determine the most appropriate values of parameters for adjusting the relative effect of the positive and negative training samples, a Genetic algorithm is used as an optimization technique. The purposed method is used to classify the internal carotid artery Doppler signals recorded from 136 patients as well as of 55 healthy people. Our proposed method reached 97.38% classification accuracy with fivefold cross-validation (CV) technique. The classification results showed that the proposed method was effective for the classification of internal carotid artery Doppler signals.

Narrowing of the middle cerebral artery: artificial intelligence methods and comparison of transcranial color coded duplex sonography with conventional TCD

The goal of the study was to compare performances of transcranial color-coded duplex sonography (TCCS) and transcranial Doppler sonography (TCD) in the diagnosis of the middle cerebral artery (MCA) narrowing in the same population of patients using statistical and nonstatistical intelligent models for data analysis. We prospectively collected data from 179 consecutive routine digital subtraction angiography (DSA) procedures performed in 111 patients (mean age 54.17+/-14.4 years; 59 women, 52 men) who underwent TCD and TCCS examinations simultaneously. Each patient was examined independently using both ultrasound techniques, 267 M1 segments of MCA were assessed and narrowings were classified as < or =50% and >50% lumen reduction. Diagnostic performance was estimated by two statistical and two artificial neural networks (ANN) classification methods. Separate models were constructed for the TCD and TCCS sonographic data, as well as for detection of "any narrowing" and "severe narrowing" of the MCA. Input for each classifier consisted of the peak-systolic, mean and end-diastolic velocities measured with each sonographic method; the output was MCA narrowing. Arterial narrowings less or equal 50% of lumen reduction were found in 55 and >50% narrowings in 26 out of 267 arteries, as indicated by DSA. In the category of "any narrowing" the rate of correct assignment by all models was 82% to 83% for TCCS and 79% to 81% for TCD. In the diagnosis of >50% narrowing the overall classification accuracy remained in the range of 89% to 90% for TCCS data and 90% to 91% for TCD data. For the diagnosis of any narrowing, the sensitivity of the TCCS was significantly higher than that of the TCD, while for diagnosis of >50% MCA narrowing, sensitivity of the TCCS was similar to sensitivity of the TCD. Our study showed that TCCS outperforms conventional TCD in detection of < or =50% MCA narrowing, whereas no significant difference in accuracy between both methods was found in the diagnosis of >50% MCA narrowing. (E-mail: jaroslaw.krejza@uphs.upenn.edu).

A multivariate density estimator for contrast agent injection monitoring using a Bayesian sparse kernel approach

The administration of intravenous contrast media during CT examinations is routine, but carries with it a risk of extravasation. Hence, we define an injection state to be either intravenous or extravasated. With a new Doppler ultrasound monitoring technique, we propose a method for estimating the probability of an injection state during the various stages of an examination. A smoothed time-frequency representation of the Doppler signal is used to analyze at which frequencies there is the largest difference in response between signals from intravenous and extravasated injections. A vector of response values based on this analysis forms this study's feature space. A Relevance Vector Machine is used to estimate the probability density for a particular injection state. We present preliminary results (n=5) showing the time-frequency representation of the Doppler ultrasound signal, the frequency analysis and the density estimation.