Artificial Intelligence Mediated Neuroradiology

Abstract

Computer technology is rapidly evolving healthcare. Improvements in image processing and subsequent computer programming have led to the development of highly accurate neural networks capable of computer-assisted diagnoses (CAD). CAD has improved clinician diagnostic accuracy for skin cancer, breast cancer, tuberculosis, and other pathologies.1,2 While accurate CAD for neurological diseases have yet to be achieved, Titano et al3 applied neural networks to triage cranial computed tomography (CT) images. Prior to reviewing their publication, we will introduce the concepts behind machine learning and neural networks. Machine learning is the science that “focuses on the question of how to get computers to program themselves.”4 A popular method of machine learning is neural network deep learning, a simple example is shown in the Figure. While a human brain comprises neurons, a neural network is comprised of “nodes.” In the human brain, neurons fire in a dichotomous fashion (on or off); whereas, network nodes contain a number (or proportion) ranging from 0 to 1. The input nodes are derived from the input image; usually each singular input node acts as a numerical representation of the black and white (0 to 1) intensity of each pixel in the input image. Therefore, a 10 by 10 pixelated image corresponds to a 100 nodal input layer. The output layer is often a set of nodes that each represents a numerical probability of a specific diagnosis (Figure; Tumor, Hemorrhage, and Ischemia). In between the input and output layers are a large number of nodal layers (Figure displays 1 intermediate layer). Every node in a layer is determined by the all nodes in the prior layer. If the input image is 3D, such as a cranial CT, then each input node often represents a voxel, or 3D pixel, intensity. The speed and accuracy of a network depends on the image pre-processing and the complex architecture of the intermediate nodal layers.FIGURE.: Simplified neural network.To ensure output probabilities are accurate, neural networks have to be trained on validated images. Training the network includes inputting training images, receiving the computed output diagnosis probabilities, comparing the network's predicted diagnosis with the verified diagnosis and back propagating a correction function throughout the various intermediate nodal layers to achieve an “improved network.” After the network “learns” from thousands of training images, it is validated with a novel imaging set to obtain the network's sensitivity and specificity. Rather than using neural networks for CAD, Titano et al3 applied machine learning to triage cranial images aiming to reduce time to treatment in emergent cases. With a dataset of 37 236 head CTs and their written reports, the authors trained a neural network to read CT scans and predict the level of acuity. To accomplish this, they first labeled the training data with diagnostic terms derived from the Universal Medical Language System through the use of a natural language processing (NPL) pipeline that they previously reported.4 An NPL is a program that assigns a general label to a text entry, in this case a singular diagnostic term to the CT reports. Their NPL had a previously reported 93%sensitivity and 90% specificity for identifying a critical finding in written cranial CT report.4 Each diagnostic term was designated as critical or non-critical by a physician. For example, a CT report with an NPL designated label of “venous sinus thrombosis” was critical; whereas, “bilateral sub-acute lacune” was non-critical. Using the CTs labeled as critical or non-critical by the NPL pipeline, they trained a voxel based neural network to determine the level of acuity for cranial CTs. The neural network, electronic medical record (EMR) prioritization system, and 3 physicians were tested and compared with a set of 180 novel CT images. The network had an observed sensitivity, specificity, and area under the curve (AUC) of 0.79, 0.48, and 0.73, respectively. This was far superior to the institutions EMR (accuracy 0.51, specificity 0.73, and sensitivity 0.16) and comparable to the physicians (sensitivity 0.79 and specificity 0.85). The average network time required to determine the acuity level and que a novel CT image was 134 ms. In a small simulated randomized clinical trial, the network was able to raise an alarm for an image 150 times faster than a human. Treatment outcomes for ischemic stroke, hydrocephalus, hemorrhage, and other neurosurgical emergencies largely depend on the time to treatment. Decreased emergency department to treatment time is correlated with improved patient outcomes in stroke and traumatic acute subdural hematoma patients.5,6 Titano et al3 aimed to decrease time to treatment through a neural network driven triage system. While the simulated RCT demonstrated promising system efficacy, its real-world application is questionable. The trial failed to account for emergency situations, where physicians often call radiologists directly to expedite CT review. The authors intend to conduct a multicenter RCT in the clinical environment to validate their current findings. The main limitation of the current neurological CAD is the high false positive rates; however, Titano et al3 circumvented this by redirecting the network to output image acuity rather than diagnoses prediction. They accepted a high false positive rate, knowing that a triage system requires speed and a low false negative rate. Furthermore, a triage-focused algorithm brings urgent images to human eyes; therefore, human review of incorrectly flagged images has no serious clinical cost. The same cannot be said about false positives in a CAD system, where incorrect diagnoses can lead to improper treatment. Regardless, future developments in computer hardware and software will likely increase the accuracy, speed, and clinical application of these intelligent algorithms. Disclosure The authors have no personal, financial, or institutional interest in any of the drugs, materials, or devices described in this article.

Ähnliche Arbeiten