Improving reference standards for validation of AI-based radiography

Performance changes due to differences among annotating radiologists for training data in computerized lesion detection

PURPOSE

The quality and bias of annotations by annotators (e.g., radiologists) affect the performance changes in computer-aided detection (CAD) software using machine learning. We hypothesized that the difference in the years of experience in image interpretation among radiologists contributes to annotation variability. In this study, we focused on how the performance of CAD software changes with retraining by incorporating cases annotated by radiologists with varying experience.

METHODS

We used two types of CAD software for lung nodule detection in chest computed tomography images and cerebral aneurysm detection in magnetic resonance angiography images. Twelve radiologists with different years of experience independently annotated the lesions, and the performance changes were investigated by repeating the retraining of the CAD software twice, with the addition of cases annotated by each radiologist. Additionally, we investigated the effects of retraining using integrated annotations from multiple radiologists.

RESULTS

The performance of the CAD software after retraining differed among annotating radiologists. In some cases, the performance was degraded compared to that of the initial software. Retraining using integrated annotations showed different performance trends depending on the target CAD software, notably in cerebral aneurysm detection, where the performance decreased compared to using annotations from a single radiologist.

CONCLUSIONS

Although the performance of the CAD software after retraining varied among the annotating radiologists, no direct correlation with their experience was found. The performance trends differed according to the type of CAD software used when integrated annotations from multiple radiologists were used.

Multicentre external validation of a commercial artificial intelligence software to analyse chest radiographs in health screening environments with low disease prevalence

OBJECTIVES

To externally validate the performance of a commercial AI software program for interpreting CXRs in a large, consecutive, real-world cohort from primary healthcare centres.

METHODS

A total of 3047 CXRs were collected from two primary healthcare centres, characterised by low disease prevalence, between January and December 2018. All CXRs were labelled as normal or abnormal according to CT findings. Four radiology residents read all CXRs twice with and without AI assistance. The performances of the AI and readers with and without AI assistance were measured in terms of area under the receiver operating characteristic curve (AUROC), sensitivity, and specificity.

RESULTS

The prevalence of clinically significant lesions was 2.2% (68 of 3047). The AUROC, sensitivity, and specificity of the AI were 0.648 (95% confidence interval [CI] 0.630-0.665), 35.3% (CI, 24.7-47.8), and 94.2% (CI, 93.3-95.0), respectively. AI detected 12 of 41 pneumonia, 3 of 5 tuberculosis, and 9 of 22 tumours. AI-undetected lesions tended to be smaller than true-positive lesions. The readers' AUROCs ranged from 0.534-0.676 without AI and 0.571-0.688 with AI (all p values < 0.05). For all readers, the mean reading time was 2.96-10.27 s longer with AI assistance (all p values < 0.05).

CONCLUSIONS

The performance of commercial AI in these high-volume, low-prevalence settings was poorer than expected, although it modestly boosted the performance of less-experienced readers. The technical prowess of AI demonstrated in experimental settings and approved by regulatory bodies may not directly translate to real-world practice, especially where the demand for AI assistance is highest.

KEY POINTS

• This study shows the limited applicability of commercial AI software for detecting abnormalities in CXRs in a health screening population. • When using AI software in a specific clinical setting that differs from the training setting, it is necessary to adjust the threshold or perform additional training with such data that reflects this environment well. • Prospective test accuracy studies, randomised controlled trials, or cohort studies are needed to examine AI software to be implemented in real clinical practice.

Critical Appraisal of Artificial Intelligence-Enabled Imaging Tools Using the Levels of Evidence System

Clinical adoption of an artificial intelligence-enabled imaging tool requires critical appraisal of its life cycle from development to implementation by using a systematic, standardized, and objective approach that can verify both its technical and clinical efficacy. Toward this concerted effort, the ASFNR/ASNR Artificial Intelligence Workshop Technology Working Group is proposing a hierarchal evaluation system based on the quality, type, and amount of scientific evidence that the artificial intelligence-enabled tool can demonstrate for each component of its life cycle. The current proposal is modeled after the levels of evidence in medicine, with the uppermost level of the hierarchy showing the strongest evidence for potential impact on patient care and health care outcomes. The intended goal of establishing an evidence-based evaluation system is to encourage transparency, foster an understanding of the creation of artificial intelligence tools and the artificial intelligence decision-making process, and to report the relevant data on the efficacy of artificial intelligence tools that are developed. The proposed system is an essential step in working toward a more formalized, clinically validated, and regulated framework for the safe and effective deployment of artificial intelligence imaging applications that will be used in clinical practice.

Repeatability of radiographic assessments for feline hip dysplasia suggest consensus scores in radiology are more uncertain than commonly assumed

Variation in the diagnostic interpretation of radiographs is a well-recognised problem in human and veterinary medicine. One common solution is to create a 'consensus' score based on a majority or unanimous decision from multiple observers. While consensus approaches are generally assumed to improve diagnostic repeatability, the extent to which consensus scores are themselves repeatable has rarely been examined. Here we use repeated assessments by three radiologists of 196 hip radiographs from 98 cats within a health-screening programme to examine intra-observer, inter-observer, majority-consensus and unanimous-consensus repeatability scores for feline hip dysplasia. In line with other studies, intra-observer and inter-observer repeatability was moderate (63-71%), and related to the reference assessment and time taken to reach a decision. Consensus scores did show reduced variation between assessments compared to individuals, but consensus repeatability was far from perfect. Only 75% of majority consensus scores were in agreement between assessments, and based on Bayesian multinomial modelling we estimate that unanimous consensus scores can have repeatabilities as low as 83%. These results clearly show that consensus scores in radiology can have large uncertainties, and that future studies in both human and veterinary medicine need to include consensus-uncertainty estimates if we are to properly interpret radiological diagnoses and the extent to which consensus scores improve diagnostic accuracy.