Spatial localization accuracy of radiologists in free-response studies: Inferring perceptual FROC curves from mark-rating data

Joint Localization and Classification of Breast Cancer in B-Mode Ultrasound Imaging via Collaborative Learning with Elastography

Convolutional neural networks (CNNs) have been successfully applied in the computer-aided ultrasound diagnosis for breast cancer. Up to now, several CNN-based methods have been proposed. However, most of them consider tumor localization and classification as two separate steps, rather than performing them simultaneously. Besides, they suffer from the limited diagnosis information in the B-mode ultrasound (BUS) images. In this study, we develop a novel network ResNet-GAP that incorporates both localization and classification into a unified procedure. To enhance the performance of ResNet-GAP, we leverage stiffness information in the elastography ultrasound (EUS) modality by collaborative learning in the training stage. Specifically, a dual-channel ResNet-GAP is developed, one channel for BUS and the other for EUS. In each channel, multiple class activity maps (CAMs) are generated using a series of convolutional kernels of different sizes. The multi-scale consistency of the CAMs in both channels are further considered in network optimization. Experiments on 264 patients in this study show that the newly developed ResNet-GAP achieves an accuracy of 88.6%, a sensitivity of 95.3%, a specificity of 84.6%, and an AUC of 93.6% on the classification task, and a 1.0NLF of 87.9% on the localization task, which is better than some state-of-the-art approaches.

Computerized detection of breast lesions using deformable part models in ultrasound images

Ultrasound imaging is considered an important complementary technique for the screening of dense breasts. Detection of lesions at an early stage is a key step in which computerized lesion detection systems could play an important role in the analysis of US images. In this article, we propose adaptation of a generic object detection technique, deformable part models, to detect lesions in breast US images. The data set used in this study included 326 images, all from different patients (54 malignant lesions, 109 benign lesions and 163 healthy breasts). In terms of lesion detection, our proposal outperformed some of the most relevant approaches described in the literature; we obtained a sensitivity of 86% with 0.28 false-positive detection per image and an Az value of 0.975. In the detection of malignant lesions, our proposed approached had an Az value of 0.93 and a sensitivity of 78% at a 1.15 false-positive detections per image.

Towards a framework for analysis of eye-tracking studies in the three dimensional environment: a study of visual search by experienced readers of endoluminal CT colonography

OBJECTIVE

Eye tracking in three dimensions is novel, but established descriptors derived from two-dimensional (2D) studies are not transferable. We aimed to develop metrics suitable for statistical comparison of eye-tracking data obtained from readers of three-dimensional (3D) "virtual" medical imaging, using CT colonography (CTC) as a typical example.

METHODS

Ten experienced radiologists were eye tracked while observing eight 3D endoluminal CTC videos. Subsequently, we developed metrics that described their visual search patterns based on concepts derived from 2D gaze studies. Statistical methods were developed to allow analysis of the metrics.

RESULTS

Eye tracking was possible for all readers. Visual dwell on the moving region of interest (ROI) was defined as pursuit of the moving object across multiple frames. Using this concept of pursuit, five categories of metrics were defined that allowed characterization of reader gaze behaviour. These were time to first pursuit, identification and assessment time, pursuit duration, ROI size and pursuit frequency. Additional subcategories allowed us to further characterize visual search between readers in the test population.

CONCLUSION

We propose metrics for the characterization of visual search of 3D moving medical images. These metrics can be used to compare readers' visual search patterns and provide a reproducible framework for the analysis of gaze tracking in the 3D environment.

ADVANCES IN KNOWLEDGE

This article describes a novel set of metrics that can be used to describe gaze behaviour when eye tracking readers during interpretation of 3D medical images. These metrics build on those established for 2D eye tracking and are applicable to increasingly common 3D medical image displays.

A free-response evaluation determining value in the computed tomography attenuation correction image for revealing pulmonary incidental findings: a phantom study

RATIONALE AND OBJECTIVES

The purpose of this study was to compare lesion-detection performance when interpreting computed tomography (CT) images that are acquired for attenuation correction when performing single photon emission computed tomography/computed tomography (SPECT/CT) myocardial perfusion studies. In the United Kingdom, there is a requirement that these images be interpreted; thus, it is necessary to understand observer performance on these images.

MATERIALS AND METHODS

An anthropomorphic chest phantom with inserted spherical lesions of different sizes and contrasts was scanned on five different SPECT/CT systems using site-specific CT protocols for SPECT/CT myocardial perfusion imaging. Twenty-one observers (0-4 years of CT experience) searched 26 image slices (17 abnormal, containing 1-3 lesions, and 9 normal, containing no lesions) for each CT acquisition. The observers marked and rated perceived lesions under the free-response paradigm. Four analyses were conducted using jackknife alternative free-response receiver operating characteristic (JAFROC) analysis: (1) 20-pixel acceptance radius (AR) with all 21 readers, abbreviated to 20/ALL analysis, (2) 40-pixel AR with 21 readers (40/ALL), (3) 20-pixel AR with 14 readers experienced in CT (20/EXP), and (4) 20-pixel AR with 7 readers with no CT experience (20/NOT). The significance level of the test was set so as to conservatively control the overall probability of a type I error to <0.05.

RESULTS

The mean JAFROC figure of merit (FOM) for the five CT acquisitions for the 20/ALL study were 0.602, 0.639, 0.372, 0.475, and 0.719 with a significant difference in lesion-detection performance evident between all individual treatment pairs (P < .0001) with the exception of the 1-2 pairing, which was not significant (these differed only in milliamp seconds). System 5, which had the highest performance, had the smallest slice thickness and the largest matrix size. For the other analyses, the system orderings remained unchanged, and the significance of FOM difference findings remained identical to those for 20/ALL, with one exception: for 20/EXP analysis the 1-2 difference became significant with the higher milliamp seconds superior. Improved detection performance was associated with a smaller slice thickness, increased matrix size, and, to a lesser extent, increased tube charge.

CONCLUSIONS

Protocol variations for CT-based attenuation correction (AC) in SPECT/CT imaging have a measurable impact on lesion-detection performance. The results imply that z-axis resolution and matrix size had the greatest impact on lesion detection, with a weaker but detectable dependence on the product of milliamp and seconds.

A brief history of free-response receiver operating characteristic paradigm data analysis

In the receiver operating characteristic paradigm the observer assigns a single rating to each image and the location of the perceived abnormality, if any, is ignored. In the free-response receiver operating characteristic paradigm the observer is free to mark and rate as many suspicious regions as are considered clinically reportable. Credit for a correct localization is given only if a mark is sufficiently close to an actual lesion; otherwise, the observer's mark is scored as a location-level false positive. Until fairly recently there existed no accepted method for analyzing the resulting relatively unstructured data containing random numbers of mark-rating pairs per image. This report reviews the history of work in this field, which has now spanned more than five decades. It introduces terminology used to describe the paradigm, proposed measures of performance (figures of merit), ways of visualizing the data (operating characteristics), and software for analyzing free-response receiver operating characteristic studies.

ROCView: prototype software for data collection in jackknife alternative free-response receiver operating characteristic analysis

ROCView has been developed as an image display and response capture (IDRC) solution to image display and consistent recording of reader responses in relation to the free-response receiver operating characteristic paradigm. A web-based solution to IDRC for observer response studies allows observations to be completed from any location, assuming that display performance and viewing conditions are consistent with the study being completed. The simplistic functionality of the software allows observations to be completed without supervision. ROCView can display images from multiple modalities, in a randomised order if required. Following registration, observers are prompted to begin their image evaluation. All data are recorded via mouse clicks, one to localise (mark) and one to score confidence (rate) using either an ordinal or continuous rating scale. Up to nine "mark-rating" pairs can be made per image. Unmarked images are given a default score of zero. Upon completion of the study, both true-positive and false-positive reports can be downloaded and adapted for analysis. ROCView has the potential to be a useful tool in the assessment of modality performance difference for a range of imaging methods.

On the choice of acceptance radius in free-response observer performance studies

OBJECTIVES

Choosing an acceptance radius or proximity criterion is necessary to analyse free-response receiver operating characteristic (FROC) observer performance data. This is currently subjective, with little guidance in the literature about what is an appropriate acceptance radius. We evaluated varying acceptance radii in a nodule detection task in chest radiography and suggest guidelines for determining an acceptance radius.

METHODS

80 chest radiographs were chosen, half of which contained nodules. We determined each nodule's centre. 21 radiologists read the images. We created acceptance radii bins of <5 pixels, <10 pixels, <20 pixels and onwards up to <200 and 200+ pixels. We counted lesion localisations in each bin and visually compared marks with the borders of nodules.

RESULTS

Most reader marks were tightly clustered around nodule centres, with tighter clustering for smaller than for larger nodules. At least 70% of readers' marks were placed within <10 pixels for small nodules, <20 pixels for medium nodules and <30 pixels for large nodules. Of 72 inspected marks that were less than 50 pixels from the centre of a nodule, only 1 fell outside the border of a nodule.

CONCLUSION

The acceptance radius should be based on the larger nodule sizes. For our data, an acceptance radius of 50 pixels would have captured all but 2 reader marks within the borders of a nodule, while excluding only 1 true-positive mark. The choice of an acceptance radius for FROC analysis of observer performance studies should be based on the size of larger abnormalities.

Quantifying the clinical relevance of a laboratory observer performance paradigm

OBJECTIVE

Laboratory observer performance measurements, receiver operating characteristic (ROC) and free-response ROC (FROC) differ from actual clinical interpretations in several respects, which could compromise their clinical relevance. The objective of this study was to develop a method for quantifying the clinical relevance of a laboratory paradigm and apply it to compare the ROC and FROC paradigms in a nodule detection task.

METHODS

The original prospective interpretations of 80 digital chest radiographs were classified by the truth panel as correct (C=1) or incorrect (C=0), depending on correlation with additional imaging, and the average of C was interpreted as the clinical figure of merit. FROC data were acquired for 21 radiologists and ROC data were inferred using the highest ratings. The areas under the ROC and alternative FROC curves were used as laboratory figures of merit. Bootstrap analysis was conducted to estimate conventional agreement measures between laboratory and clinical figures of merit. Also computed was a pseudovalue-based image-level correctness measure of the laboratory interpretations, whose association with C as measured by the area (rAUC) under an appropriately defined relevance ROC curve, is as a measure of the clinical relevance of a laboratory paradigm.

RESULTS

Low correlations (e.g. κ=0.244) and near chance level rAUC values (e.g. 0.598), attributable to differences between the clinical and laboratory paradigms, were observed. The absolute width of the confidence interval was 0.38 for the interparadigm differences of the conventional measures and 0.14 for the difference of the rAUCs.

CONCLUSION

The rAUC measure was consistent with the traditional measures but was more sensitive to the differences in clinical relevance. A new relevance ROC method for quantifying the clinical relevance of a laboratory paradigm is proposed.

New developments in observer performance methodology in medical imaging

A common task in medical imaging is assessing whether a new imaging system, or a variant of an existing one, is an improvement over an existing imaging technology. Imaging systems are generally quite complex, consisting of several components-for example, image acquisition hardware, image processing and display hardware and software, and image interpretation by radiologists- each of which can affect performance. Although it may appear odd to include the radiologist as a "component" of the imaging chain, because the radiologist's decision determines subsequent patient care, the effect of the human interpretation has to be included. Physical measurements such as modulation transfer function, signal-to-noise ratio, are useful for characterizing the nonhuman parts of the imaging chain under idealized and often unrealistic conditions, such as uniform background phantoms and target objects with sharp edges. Measuring the performance of the entire imaging chain, including the radiologist, and using real clinical images requires different methods that fall under the rubric of observer performance methods or "ROC" analysis, that involve collecting rating data on images. The purpose of this work is to review recent developments in this field, particularly with respect to the free-response method, where location information is also collected.

Recent developments in imaging system assessment methodology, FROC analysis and the search model

A frequent problem in imaging is assessing whether a new imaging system is an improvement over an existing standard. Observer performance methods, in particular the receiver operating characteristic (ROC) paradigm, are widely used in this context. In ROC analysis lesion location information is not used and consequently scoring ambiguities can arise in tasks, such as nodule detection, involving finding localized lesions. This paper reviews progress in the free-response ROC (FROC) paradigm in which the observer marks and rates suspicious regions and the location information is used to determine whether lesions were correctly localized. Reviewed are FROC data analysis, a search-model for simulating FROC data, predictions of the model and a method for estimating the parameters. The search model parameters are physically meaningful quantities that can guide system optimization.

The effect of gadoxetic acid enhancement on lesion detection and characterisation using T₂ weighted imaging and diffusion weighted imaging of the liver

OBJECTIVES

To evaluate the effect of gadoxetic acid enhancement on the detection and characterisation of focal hepatic lesions on T(2) weighted and diffusion weighted (DW) images.

METHODS

A total of 63 consecutive patients underwent T(2) weighted and DW imaging before and after gadoxetic acid enhancement. Two blinded readers independently identified all of the focal lesions using a five-point confidence scale and characterised each lesion using a three-point scale: 1, non-solid; 2, indeterminate; and 3, solid. For both T(2) weighted and DW imaging, the accuracies for detecting focal lesions were compared using the free-response receiver operating characteristic analysis; the accuracies for lesion characterisation were compared using the McNemar test between non-enhanced and gadoxetic acid-enhanced image sets. For hepatic lesions ≥ 1 cm, the lesion-to-liver contrast-to-noise ratio (CNR) and the apparent diffusion coefficient (ADC) were compared in the non-enhanced and enhanced image sets using the generalised estimating equations.

RESULTS

For both T(2) weighted and DW images, the accuracies for detecting focal lesions (p ≥ 0.52) and those for lesion characterisation (p ≥ 0.63) did not differ significantly between the non-enhanced and enhanced image sets. The lesion-to-liver CNR was significantly higher on enhanced DW images than on non-enhanced DW images (p=0.02), although the difference was not significant for T(2) weighted imaging (p=0.65). The mean ADC values of lesions did not differ significantly on enhanced and non-enhanced DW imaging (p=0.75).

CONCLUSION

The acquisition of T(2) weighted and DW images after administration of gadoxetic acid has no significant effect on the detection or characterisation of focal hepatic lesions, although it improves the lesion-to-liver CNR on DW images.

A status report on free-response analysis

The purpose of this paper is to summarise recent progress in free-response receiver operating characteristic (FROC) methodology. These are: (1) jackknife alternative FROC analysis including recent extensions and alternative methods; (2) the search-model simulator that enables validation and objective comparison of different methods of analysing the data; (3) case-based analysis that has the potential of greater clinical relevance than conventional free-response analysis; (4) a method for collectively analysing the multiple lesion types in an image (e.g. microcalcifications, masses and architectural distortions); (5) a method for sample-size estimation for FROC studies; and (6) a method for determining an objective proximity criterion, namely how close must a mark be to a true lesion in order to credit the observer for a true localisation. FROC analysis is being increasingly used to evaluate the imaging systems and understanding of recent progress should help researchers conduct better FROC studies.

Consistency of methods for analysing location-specific data

Although the receiver operating characteristic (ROC) method is the acknowledged gold-standard for imaging system assessment, it ignores localisation information and differentiation between multiple abnormalities per case. As the free-response ROC (FROC) method uses localisation information and more closely resembles the clinical reporting process, it is being increasingly used. A number of methods have been proposed to analyse the data that result from an FROC study: jackknife alternative FROC (JAFROC) and a variant termed JAFROC1, initial detection and candidate analysis (IDCA) and ROC analysis via the reduction of the multiple ratings on a case to a single rating. The focus of this paper was to compare JAFROC1, IDCA and the ROC analysis methods using a clinical FROC human data set. All methods agreed on the ordering of the modalities and all yielded statistically significant differences of the figures-of-merit, i.e. p < 0.05. Both IDCA and JAFROC1 yielded much smaller p-values than ROC. The results are consistent with a recent simulation-based validation study comparing these and other methods. In conclusion, IDCA or JAFROC1 analysis of FROC human data may be superior at detecting modality differences than ROC analysis.

Agreement of the order of overall performance levels under different reading paradigms

RATIONALE AND OBJECTIVES

To investigate consistency of the orders of performance levels when interpreting mammograms under three different reading paradigms.

MATERIALS AND METHODS

We performed a retrospective observer study in which nine experienced radiologists rated an enriched set of mammography examinations that they personally had read in the clinic ("individualized") mixed with a set that none of them had read in the clinic ("common set"). Examinations were interpreted under three different reading paradigms: binary using screening Breast Imaging Reporting and Data System (BI-RADS), receiver-operating characteristic (ROC), and free-response ROC (FROC). The performance in discriminating between cancer and noncancer findings under each of the paradigms was summarized using Youden's index/2+0.5 (Binary), nonparameteric area under the ROC curve (AUC), and an overall FROC index (JAFROC-2). Pearson correlation coefficients were then computed to assess consistency in the ordering of observers' performance levels. Statistical significance of the computed correlation coefficients was assessed using bootstrap confidence intervals obtained by resampling sets of examination-specific observations.

RESULTS

All but one of the computed pair-wise correlation coefficients were larger than 0.66 and were significantly different from zero. The correlation between the overall performance measures under the Binary and ROC paradigms was the lowest (0.43) and was not significantly different from zero (95% confidence interval -0.078 to 0.733).

CONCLUSION

The use of different evaluation paradigms in the laboratory tends to lead to consistent ordering of the overall performance levels of observers. However, one should recognize that conceptually similar performance indexes resulting from different paradigms often measure different performance characteristics and thus disagreements are not only possible but frequently quite natural.

Performance assessments of diagnostic systems under the FROC paradigm: experimental, analytical, and results interpretation issues

As use of free response receiver-operating characteristic (FROC) curves gains more acceptance for quantitatively assessing the performance of diagnostic systems, it is important that the experimentalist understands the possible role of this approach as one of the experimental design paradigms that are available to him or her among all other approaches as well as some of the issues associated with FROC type studies. In a number of experimental scenarios, the FROC paradigm and associated analytical tools have theoretical and practical advantages over both the binary and the ROC approaches to performance assessments of diagnostic systems, but it also has some limitations related to experimental design, data analyses, clinical relevance, and complexity in the interpretation of the results. These issues are rarely discussed and are the focus of this work.

Operating characteristics predicted by models for diagnostic tasks involving lesion localization

In 1996 Swensson published an observer model that predicted receiver operating characteristic (ROC), localization ROC (LROC), free-response ROC (FROC) and alternative FROC (AFROC) curves, thereby achieving "unification" of different observer performance paradigms. More recently a model termed initial detection and candidate analysis (IDCA) has been proposed for fitting computer aided detection (CAD) generated FROC data, and recently a search model for human observer FROC data has been proposed. The purpose of this study was to derive IDCA and the search model based expressions for operating characteristics, and to compare the predictions to the Swensson model. For three out of four mammography CAD data sets all models yielded good fits in the high-confidence region, i.e., near the lower end of the plots. The search model and IDCA tended to better fit the data in the low-confidence region, i.e., near the upper end of the plots, particularly for FROC curves for which the Swensson model predictions departed markedly from the data. For one data set none of the models yielded satisfactory fits. A unique characteristic of search model and IDCA predicted operating characteristics is that the operating point is not allowed to move continuously to the lowest confidence limit of the corresponding Swensson model curves. This prediction is actually observed in the CAD raw data and it is the primary reason for the poor FROC fits of the Swensson model in the low-confidence region.

Evaluating computer-aided detection algorithms

Computer-aided detection (CAD) has been attracting extensive research interest during the last two decades. It is recognized that the full potential of CAD can only be realized by improving the performance and robustness of CAD algorithms and this requires good evaluation methodology that would permit CAD designers to optimize their algorithms. Free-response receiver operating characteristic (FROC) curves are widely used to assess CAD performance, however, evaluation rarely proceeds beyond determination of lesion localization fraction (sensitivity) at an arbitrarily selected value of nonlesion localizations (false marks) per image. This work describes a FROC curve fitting procedure that uses a recent model of visual search that serves as a framework for the free-response task. A maximum likelihood procedure for estimating the parameters of the model from free-response data and fitting CAD generated FROC curves was implemented. Procedures were implemented to estimate two figures of merit and associated statistics such as 95% confidence intervals and goodness of fit. One of the figures of merit does not require the arbitrary specification of an operating point at which to evaluate CAD performance. For comparison a related method termed initial detection and candidate analysis was also implemented that is applicable when all suspicious regions are reported. The two methods were tested on seven mammography CAD data sets and both yielded good to excellent fits. The search model approach has the advantage that it can potentially be applied to radiologist generated free-response data where not all suspicious regions are reported, only the ones that are deemed sufficiently suspicious to warrant clinical follow-up. This work represents the first practical application of the search model to an important evaluation problem in diagnostic radiology. Software based on this work is expected to benefit CAD developers working in diverse areas of medical imaging.