Applying Machine Vision Techniques to Neuroimaging

Abstract

Machine vision, the branch of artificial intelligence dealing with the automated extraction of information from images, has recently received significant media attention. Popular applications of machine vision include self-driving cars and automatic facial recognition for security footage. However, few applications exist in either the clinical setting or in neurosurgery. A recent study published in Nature Medicine by Titano et al1 aims to change that.FIGURE.: Results from Titano et al 1 of a randomized vs CNN-prioritized radiologist work flow. A, box plots showing 25th, 50th, and 75th percentile times for individual imaging processing, demonstrating a several order-of-magnitude speed advantage for machine interpretation. B, Distribution of work-queue positions of critical vs noncritical studies in a randomized vs CNN-prioritized workflow, demonstrating earlier viewing of critical images under CNN-prioritization. C, Representation schema showing how CNN-prioritization under high-sensitivity settings would shift queue positions of potentially critical images to be viewed earlier, even with a high false-positive rate. Reprinted by permission from Springer Nature: Springer Nature, Nature Medicine, Automated deep-neural-network surveillance of cranial images for acute neurologic events, Joseph J. Titano et al,1 Copyright Springer Nature, 2018.This study involved training a 3-dimensional convolutional neural network (3D-CNN) to recognize potentially urgent or acute imaging findings on a series of noncontrasted head computed tomography (NCHCT) scans. First, 37 236 NCHCTs and 96 303 radiology reports were obtained from historical patient records. Using a natural language processing (NLP) algorithm, diagnostic labels were automatically extracted from the reports. Next, the trained NLP was used to analyze the text reports of the 37 236 NCHCTs and the researchers mapped those diagnostic labels to a given acuity level (critical vs noncritical). The 3D-CNN was subsequently trained to map the actual imaging data of the NCHCTs to diagnoses obtained from the NLP. To test the results of this, a small double-blinded randomized trial was performed where 180 NCHCTs were assigned to radiologists in either a random order or a 3D-CNN-prioritized order. In identifying critical vs noncritical values, the 3D-CNN interpretation had a sensitivity of 0.79 and a specificity of 0.48 (with gold-standard labels obtained from manual chart review including subsequent information of the patient's clinical course), compared to human interpretation, which had a sensitivity and specificity of 0.79 and 0.85, respectively. Unsurprisingly, the machine vision algorithm was much faster than humans at interpreting images and raising alarms (1.2 s vs 177 s on average, P < .0001). Moreover, authors were able to show that a 3D-CNN-prioriztied work flow was able to consistently place critical studies earlier in the work flow (P = .01), potentially shaving minutes off the alert times for critical images in actual clinical practice with long work queues (Figure). This study demonstrates that it is feasible to create a machine vision algorithm for aid in interpretation of urgent neurological imaging and that it may prove useful in reducing alert times to clinicians for critical findings. However, these algorithms are characterized by a high false-positive rate, have little to no ability to discuss and interface with clinicians, and still require human radiologists for confirmation and translation of results to function effectively in a clinical setting. While computer-assisted radiology interpretation may be closer to becoming a reality, artificial intelligence is only an adjunct to real human intelligence…as it should be. Disclosure The authors have no personal, financial, or institutional interest in any of the drugs, materials, or devices described in this article.

Ähnliche Arbeiten