An analysis of perceptual errors in reading mammograms using quasi-local spatial frequency spectra.

In this pilot study the authors examined areas on a mammogram that attracted the visual attention of experienced mammographers and mammography fellows, as well as areas that were reported to contain a malignant lesion, and, based on their spatial frequency spectrum, they characterized these areas by the type of decision outcome that they yielded: true-positives (TP), false-positives (FP), true-negatives (TN), and false-negatives (FN). Five 2-view (craniocaudal and medial-lateral oblique) mammogram cases were examined by 8 experienced observers, and the eye position of the observers was tracked. The observers were asked to report the location and nature of any malignant lesions present in the case. The authors analyzed each area in which either the observer made a decision or in which the observer had prolonged (>1,000 ms) visual dwell using wavelet packets, and characterized these areas in terms of the energy contents of each spatial frequency band. It was shown that each decision outcome is characterized by a specific profile in the spatial frequency domain, and that these profiles are significantly different from one another. As a consequence of these differences, the profiles can be used to determine which type of decision a given observer will make when examining the area. Computer-assisted perception correctly predicted up to 64% of the TPs made by the observers, 77% of the FPs, and 70% of the TNs.

Blinded review of retrospectively visible unreported breast cancers: an eye-position analysis.

PURPOSE: To determine whether unreported retrospectively identified cancers on mammograms receive prolonged visual attention and can be reliably detected in a blinded review. MATERIALS AND METHODS: Four experienced mammographers performed a blinded review of a test set of 20 retrospective cases where the cancer was not detected until the next mammographic evaluation, 10 prospective cases where the cancer was initially detected, and 10 cancer-free cases. Two views were digitized and displayed on a workstation. The experiment consisted of an initial impression, during which eye position was monitored, and a final impression, during which viewers zoomed on regions of interest and localized suspicious lesions. Eye-position data were analyzed to determine whether retrospectively visible cancers attracted attention to the same degree as prospectively visible cancers. The initial impression used 1,000 msec as the eye-fixation dwell criterion for detecting a lesion. RESULTS: Initially, 70% of retrospective cancers and 50% of prospective cancers did not attract prolonged visual attention. In prospective cases, detailed examination significantly improved the mean receiver operating characteristic area, from.73 to.88 (P <.01), but in retrospective cases, the mean receiver operating characteristic area barely increased, from.60 to.68, due to a high true-positive-to-false-positive ratio. CONCLUSION: At blinded review, detection of retrospectively visible cancers was significantly inferior to that of prospective cancers. It cannot be assumed that retrospectively identified cancers are intrinsically detectable, because they do not draw prolonged visual attention during visual search for breast cancers.

How experience and training influence mammography expertise.

RATIONALE AND OBJECTIVES: The authors evaluated the influence of perceptual and cognitive skills in mammography detection and interpretation by testing three groups representing different levels of mammography expertise in terms of experience, training, and talent with a mammography screening-diagnostic task. MATERIALS AND METHODS: One hundred fifty mammograms, composed of unilateral cranial-caudal and mediolateral oblique views, were displayed in pairs on a digital workstation to 19 radiology residents, three experienced mammographers, and nine mammography technologists. One-third of the mammograms showed malignant lesions; two-thirds were malignancy-free. Observers interacted with the display to indicate whether each image contained no malignant lesions or suspicious lesions indicating malignancy. Decision time was measured as the lesions were localized, classified, and rated for decision confidence. RESULTS: Compared with performance of experts, alternative free response operating characteristic performance for residents was significantly lower and equivalent to that of technologists. Analysis of overall performance showed that, as level of expertise decreased, false-positive results exerted a greater effect on overall decision accuracy over the time course of image perception. This defines the decision speed-accuracy relationship that characterizes mammography expertise. CONCLUSION: Differences in resident performance resulted primarily from lack of perceptual-learning experience during mammography training, which limited object recognition skills and made it difficult to determine differences between malignant lesions, benign lesions, and normal image perturbations. A proposed solution is systematic mentor-guided training that links image perception to feedback about the reasons underlying decision making.

A perceptually tempered display for digital mammograms.

The cathode ray tube of a workstation for use with digital mammograms was calibrated with a photometer to produce an input-output characteristic curve similar to the perceptually linear curve defined by a current display standard. Then, a test pattern consisting of bars of increasing intensity containing disks of decreasing contrast was used by an observer to estimate the minimal detectable contrast (MDC) at different levels of display luminance. The MDC was modeled by a parabola. The shape of the parabola was determined by the observer's perceptual responses, and the range was determined by the maximum and minimum pixel values of the breast parenchyma. As each mammogram was displayed, the contour of the breast was automatically found and pixels within the breast image were sampled to determine the pixel values that were used to compute the maximum and minimum pixel values. The parabola was integrated to determine the look-up table for the initial MDC-tempered display of the mammogram. Preliminary observer performance tests showed no significant differences in the accuracy and speed of three radiologists who read a set of mammograms when the MDC-tempered display was compared with the perceptually linear display.

Nature of expertise in searching mammograms for breast masses.

RATIONALE AND OBJECTIVES: The authors investigated how training and experience affect the performance of observers searching mammograms for breast masses. METHODS: Eye positions of mammographers, mammography technologists, mammography residents, and laypersons were compared to scan paths generated by a simulated scanner as each searched nine two-view digital mammogram pairs for breast masses. RESULTS: Analysis of time-to-hit data revealed that mammographers and mammography technologists with the most extensive training and experience had the fastest search times in the detection and confirmation of a breast mass on two views. Scanning patterns of less-experienced mammography residents were less efficient due to wider dispersion of visual attention between potential breast masses and perturbations in breast parenchyma. Because laypersons lacked both training and experience in mammography, bright blobs in the breast image were considered to be intuitively valid target candidates and these features distracted the search by capturing visual attention. CONCLUSION: Experience reading normal and abnormal mammograms plays a critical role in training radiologists. Experience combined with training provides the basis for generating efficient visual search strategies and developing distinctive conceptual criteria for perceptual differentiation and interpretation of true breast masses from image artifacts and structured noise that mimics breast abnormalities.

Perceptual enhancement of tumor targets in chest X-ray images.

Under tachistoscopic viewing conditions, precuing the location of potential lung tumor targets in chest X-ray images was less effective than precuing followed by bounding the region of interest (ROI) with a circle directly on the image. Detection performance increased as the image was systematically masked so that its size approximated that of the circled ROI. When viewing time was extended to allow shifts in eye position, circling the ROI was found to restrict the dispersion of fixations and increase the accuracy of fixating the target tumor. When targets were placed outside the ROI, the circle inhibited their detection relative to detection of targets inside the circled region. These findings suggest that cuing by circling restricts target detection to the ROI, and by doing so reduces the interfering effects of outside distractors that complete with the target for attention.

Sodium nuclear magnetic resonance imaging of neuroblastoma in the nude mouse.

The average concentration of sodium is known to be elevated in some tumors relative to normal tissues, and necrosis is suspected of being a possible cause. We have performed in vivo sodium-23 magnetic resonance imaging (MRI) of IMR-5 neuroblastoma in the athymic nude mouse on a 1.9-Tesla, small-bore animal imaging system. We compared the sodium images with histologic analysis for necrosis and with proton images, chemical measurements of water and blood content, and sodium and potassium concentrations. We found that the sodium concentrations determined by MRI were proportional to the fraction of the tumor tissue that was necrotic. Correlation coefficients varied from 0.65 to 0.78, depending upon how the data were selected. With further refinement it is possible that the sodium concentration measurements determined noninvasively by MRI may have applications as part of clinical diagnosis and staging of soft tissue tumors.

Computer-displayed eye position as a visual aid to pulmonary nodule interpretation.

Approximately 30% of nodules are missed during the initial reading of chest radiographs. Eye-position recordings have shown that most nodules that are missed receive prolonged visual attention. A computer algorithm was developed that uses eye-position and gaze-duration times to identify locations on the chest image likely to contain missed nodules. These locations are highlighted on the displayed image to give visual feedback. The current study tested whether visual feedback was an effective aid to nodule detection. Six radiology residents searched 40 chest images for nodules while their eye-position and gaze-duration times were recorded. Half received displayed visual feedback and half were given a second view without feedback. Two months later the two groups returned and viewed the images in the opposite condition to counterbalance for possible practice effects. Performance of readers who were given feedback showed an average of 16% improvement as measured by the alternative free response operating characteristic (AFROC) curve area, A1. Performance of the same readers given a second look without feedback did not improve.

Searching for lung nodules. Visual dwell indicates locations of false-positive and false-negative decisions.

Eye position recordings made while radiologists searched chest images for lung nodules showed that regions falsely reported positive or suspicious received prolonged visual attention. Correlation of regional fixation dwell time with independent ratings of image features indicated that more than 90% of false-positive decisions were caused by some perturbation in the image that aroused the suspicion of the viewer. The remainder apparently arose from within the viewer. Most missed nodules (false-negative reports) also received prolonged visual attention, implying an active decision not to perceive a nodule. The data are interpreted to show that roughly one task-related decision is made during each second of scanning a radiograph. This departs from the central assumption of the traditional signal-detection model based upon one decision per image.

The diagnosis of pulmonary nodules: comparison between standard and inverse digitized images and conventional chest radiographs.

We compared plain chest radiographs, standard (bones white) digitized images, and inverse-intensity (bones black) images to determine their ability to identify pathologically confirmed malignant pulmonary nodules. The images were digitized by using a photo-optical laser scanner and were displayed on a 1024 x 1024 x 8 bit system capable of operator-controlled magnification (2x or 4x) and nonlinear (logarithmic/exponential) contrast transformation in both standard and inverse-intensity modes. Receiver-operator curve analysis was used to study the detection performance of six observers who viewed 40 images obtained in 15 normal subjects and 25 abnormal subjects. There was no statistically significant difference in the area under the ROC curve between the standard digital images and the plain chest radiographs. However, ROC areas were significantly greater (p less than or equal to .05) for inverse-intensity digital images when compared with either standard-intensity digital images or plain chest radiographs. These results suggest that inverse-intensity images may have some advantages in the detection of pulmonary nodules.

Dextran-magnetite: a contrast agent for sodium-23 MRI.

Dextran-magnetite, a superparamagnetic compound, is a powerful relaxation reagent for sodium. Administered intravenously, it is confined mainly to the vasculature and eliminates the signal from plasma sodium, a significant component of the tissue sodium signal. Applications of dextran-magnetite for in vivo sodium imaging of a normal rat and rats with a tumor and experimentally induced peripheral edema are shown. Our results indicate that dextran-magnetite may be useful for improving tumor detection and for imaging of edema. To our knowledge, this is the first report of a contrast agent suitable for sodium magnetic resonance imaging.

Diagnostic accuracy of four assays of prostatic acid phosphatase. Comparison using receiver operating characteristic curve analysis.

We compared the diagnostic characteristics of four assays of prostatic acid phosphatase by two methods. In the first analysis, receiver operating characteristic (ROC) curves were developed and the areas under the curves were calculated. In the second analysis, the sensitivity of each assay was compared with the sensitivities of the other assays with specificities set at equal levels. One or more assays were performed in 1,298 patients; 141 underwent all four assays. Pairwise comparisons of the areas under the ROC curves showed no significant differences. No differences were found in the sensitivities of the four assays when the upper limits of normal were selected to provide equal specificities. We conclude that there is little difference in diagnostic accuracy among the prostatic acid phosphatase assays. Our findings contrast with previous studies that used only one upper limit of normal and that found some assays to be superior to others. Receiver operating characteristic curve analysis corrects for the bias introduced by the choice of an upper limit of normal.

The effects of kilovoltage on diagnostic accuracy in chest radiography.

Chest examinations, consisting of three posteroanterior chest radiographs obtained at 75,120, and 140 kVp, were performed on patients with clinically suspected pulmonary disease. The 66 sets of radiographs included 23 lung cancers, 13 other clinically relevant abnormalities, and 27 normals. All findings were verified by clinical or radiologic follow-up except for three cases that could not be verified and were excluded from later analysis. Six radiologists viewed the series of 198 radiographs one at a time in a randomized sequence. They scored each radiograph and each suspected abnormality with a confidence rating from 1 (normal) to 4 (definitely abnormal). Based on these scores receiver operating characteristic (ROC) curves were plotted for each set of kVp settings. The results indicated that higher kVp values improved diagnostic accuracy. Analysis of the ROC curves showed statistically significant differences between 140 kVp and both 120 kVp and 75 kVp, but not between the two lower kVp values.

Predictive value and threshold detectability of lung tumors.

The visibility limit of soft tissue nodules in the lungs is 3 mm on a chest radiography, yet lung cancers are rarely detected until the tumors are 8-10 mm in size. This paper presents data on the size distribution of nodular lung cancers and noncancerous nodular noise on chest radiographs. These data are analyzed in decision-making terms and show that when the size is 8-10 mm, observers can separate true cancers from noise that mimics cancer, with a probability of being correct, or a predictive value, of 90% or more.

An analysis of perceptual and cognitive factors in radiographic interpretation.

Detection of small tumors in chest x-ray films was studied under tachistoscopic viewing conditions designed to stimulate single fixations varying in dwell time, found in free search. Two questions were asked: How much dwell time is required to detect a nodule (experiment 1)? How effective are peripheral inputs in the detection of a nodule (experiment 2)? Our findings indicate that a dwell time of 300 ms was sufficient to detect 85% of the nodules when they were viewed directly. Detection accuracy was reduced by one-half when the tumor was located 5 degrees from the axis of gaze. Taken together with data from eye movement experiments, these results provide useful estimates of perceptual and cognitive parameters of visual search.

Contrast gradient and the detection of lung nodules.

The effects of edge gradient and blur on the detection pulmonary nodules were studied, using a test series of 175 artificial nodules on 30 chest films. The nodules were of homogeneous texture, about 1 cm in diameter and had a mean contrast of .12. Six observers viewed the film set, scoring both location and confidence. The sensitivity or the probability of a true-positive response was linearly related to the blur and a modified ROC analysis resulted in a family of ROC curves with the index of detectability monotonically decreasing with increased blur.

Visual scanning, pattern recognition and decision-making in pulmonary nodule detection.

Eye movements were recorded while four subjects searched a set of 60 films, 24 normal and 36 abnormal for pulmonary nodules. Error rates, scanning patterns and the dwell time of fixation clusters on normal and nodule-containing areas of the film were studied. Using the assumption that prolonged dwell time indicates intensive processing of visual data, a model was developed for nodule detection that includes four steps: orientation, scanning, pattern recognition and decision-making. False-negative errors were divided into three classes: scanning errors, recognition errors and decision-making errors. Of 20 false-negative errors, 30% were considered scanning, 25% recognition and 45% decision-making.

The influence of film density on the radiologic detection of lung lesions.

Experiments were conducted on a series of chest radiographs to evaluate the effects of mean film density on reader performance. It was found that the false negative is not measurably affected, in agreement with earlier published data. It was also found, however, that the false positive rate increases with increasing film density. The causes and implications of these results are discussed.