Application of noise-reduction techniques to machine learning algorithms for breast cancer tumor identification.

The application of machine learning (ML) techniques to digitized images of biopsied cells for breast cancer diagnosis is an active area of research. We hypothesized that reducing noise in the data would lead to an increase in classification accuracies. To test this hypothesis, we first compared several classification techniques in their ability to discriminate between malignant and benign breast cancer tumors using the Wisconsin Breast Cancer Data Set and subsequently evaluated the effect of noise reduction techniques on model accuracies. We applied two noise-reduction techniques based on Principal Component Analysis - dimensionality reduction and outlier removal - to a comprehensive list of ML algorithms with different learning paradigms including Decision Trees (fine, medium, coarse), dimensionality reduction techniques (Linear Discriminant Analysis, Quadratic Discriminant Analysis, Partial Least Squares-Discriminant Analysis), logistic Regression, Bayesian techniques (Gaussian Naive, Kernel Naive), Support Vector Machines (Linear, Quadratic, Cubic, Gaussian), instance-based techniques (fine, medium, coarse, cosine, cubic, and weighted K-Nearest Neighbors), and Artificial Neural Networks. Results showed that noise removal through dimensionality reduction is most effective when using a cross-validated number of principal components, and accuracies surpassing 99% across all ML models are obtained when both noise-reduction techniques are applied sequentially. Even though such a high accuracy has been demonstrated in few instances for specific algorithms, the methodology proposed herein is the first published report demonstrating the applicability of a technique to a wide range of ML models to achieve high accuracies. We show that dimensionality reduction and outlier analysis can be used as effective approaches to improve discrimination accuracies. Also, dimensionality reduction through a cross-validated number of principal components can provide an effective framework for reducing noise in the data prior to applying a ML algorithm.

Autonomous Robotic Point-of-Care Ultrasound Imaging for Monitoring of COVID-19-Induced Pulmonary Diseases.

The COVID-19 pandemic has emerged as a serious global health crisis, with the predominant morbidity and mortality linked to pulmonary involvement. Point-of-Care ultrasound (POCUS) scanning, becoming one of the primary determinative methods for its diagnosis and staging, requires, however, close contact of healthcare workers with patients, therefore increasing the risk of infection. This work thus proposes an autonomous robotic solution that enables POCUS scanning of COVID-19 patients' lungs for diagnosis and staging. An algorithm was developed for approximating the optimal position of an ultrasound probe on a patient from prior CT scans to reach predefined lung infiltrates. In the absence of prior CT scans, a deep learning method was developed for predicting 3D landmark positions of a human ribcage given a torso surface model. The landmarks, combined with the surface model, are subsequently used for estimating optimal ultrasound probe position on the patient for imaging infiltrates. These algorithms, combined with a force-displacement profile collection methodology, enabled the system to successfully image all points of interest in a simulated experimental setup with an average accuracy of 20.6 ± 14.7 mm using prior CT scans, and 19.8 ± 16.9 mm using only ribcage landmark estimation. A study on a full torso ultrasound phantom showed that autonomously acquired ultrasound images were 100% interpretable when using force feedback with prior CT and 88% with landmark estimation, compared to 75 and 58% without force feedback, respectively. This demonstrates the preliminary feasibility of the system, and its potential for offering a solution to help mitigate the spread of COVID-19 in vulnerable environments.

A 3-D-Printed Patient-Specific Ultrasound Phantom for FAST Scan.

Ultrasound phantoms are commonly used to assess the performance of ultrasound systems and ensure their proper functionality, in addition to providing opportunities for medical training. However, Focused Assessment with Sonography for Trauma (FAST) phantoms, in particular, are prohibitively expensive and procedure specific. This work explores the use of additive manufacturing to fabricate a patient-specific, full-scale torso ultrasound phantom. Phantom geometry was derived from anonymized computed tomography scans and segments into discrete organs. The digital organs (torso, skeleton, liver, spleen) were 3-D printed and used as castable molds for producing their respective body features. These organs were integrated with artificial hemorrhages to produce a realistic training tool for FAST scans. The resulting phantom is low in cost, has a verified shelf-life of at least 1 y and was positively reviewed by a trauma and emergency radiologist for its ability to provide accurate geometric and ultrasound information.