High-Throughput Classification of Radiographs Using Deep Convolutional Neural Networks

The study aimed to determine if computer vision techniques rooted in deep learning can use a small set of radiographs to perform clinically relevant image classification with high fidelity. One thousand eight hundred eighty-five chest radiographs on 909 patients obtained between January 2013 and July 2015 at our institution were retrieved and anonymized. The source images were manually annotated as frontal or lateral and randomly divided into training, validation, and test sets. Training and validation sets were augmented to over 150,000 images using standard image manipulations. We then pre-trained a series of deep convolutional networks based on the open-source GoogLeNet with various transformations of the open-source ImageNet (non-radiology) images. These trained networks were then fine-tuned using the original and augmented radiology images. The model with highest validation accuracy was applied to our institutional test set and a publicly available set. Accuracy was assessed by using the Youden Index to set a binary cutoff for frontal or lateral classification. This retrospective study was IRB approved prior to initiation. A network pre-trained on 1.2 million greyscale ImageNet images and fine-tuned on augmented radiographs was chosen. The binary classification method correctly classified 100 % (95 % CI 99.73–100 %) of both our test set and the publicly available images. Classification was rapid, at 38 images per second. A deep convolutional neural network created using non-radiological images, and an augmented set of radiographs is effective in highly accurate classification of chest radiograph view type and is a feasible, rapid method for high-throughput annotation.     

Automated Classification of Radiology Reports to Facilitate Retrospective Study in Radiology

Retrospective research is an import tool in radiology. Identifying imaging examinations appropriate for a given research question from the unstructured radiology reports is extremely useful, but labor-intensive. Using the machine learning text-mining methods implemented in LingPipe [1], we evaluated the performance of the dynamic language model (DLM) and the Naïve Bayesian (NB) classifiers in classifying radiology reports to facilitate identification of radiological examinations for research projects. The training dataset consisted of 14,325 sentences from 11,432 radiology reports randomly selected from a database of 5,104,594 reports in all disciplines of radiology. The training sentences were categorized manually into six categories (Positive, Differential, Post Treatment, Negative, Normal, and History). A 10-fold cross-validation [2] was used to evaluate the performance of the models, which were tested in classification of radiology reports for cases of sellar or suprasellar masses and colloid cysts. The average accuracies for the DLM and NB classifiers were 88.5 % with 95 % confidence interval (CI) of 1.9 % and 85.9 % with 95 % CI of 2.0 %, respectively. The DLM performed slightly better and was used to classify 1,397 radiology reports containing the keywords “sellar or suprasellar mass”, or “colloid cyst”. The DLM model produced an accuracy of 88.2 % with 95 % CI of 2.1 % for 959 reports that contain “sellar or suprasellar mass” and an accuracy of 86.3 % with 95 % CI of 2.5 % for 437 reports of “colloid cyst”. We conclude that automated classification of radiology reports using machine learning techniques can effectively facilitate the identification of cases suitable for retrospective research.     

Instructions to authors for case reporting are limited: A review of a core journal list

Background

Case reports are frequently published in the health care literature, however advice on preparing such reports using the "instructions to authors" pages of journals is alleged to be limited. However, to our knowledge, this has not been formally evaluated. As roles of case reports may vary according to the case and the clinical specialities, one might expect the advice to authors to vary according to journal clinical grouping.

Methods

We surveyed the current advice available to authors of case reports from 'instructions to authors' pages of a core collection of 249 journals ('Hague' list). These were examined and compared for advice or recommendation on writing case reports. Of these, 163 (65%) published case reports and provided instructions on this publication type. Data were extracted on items of style and content of case reports, using a piloted data extraction form.

Results

Journals that published case reports were grouped into medical (n = 81, 50%), surgical (n = 38, 23%) and generic or multidisciplinary (n = 44, 27%) categories. There was a difference among the medical, surgical and generic or multidisciplinary journals in the maximum number of words and pages allowed but no difference in the number of figures, tables, references, authors, abstract or synopsis, indexing or key words and consent. Additionally, there was no statistically significant difference among the three different categories of journals regarding the content of the case reports.

Conclusions

Of the journals reviewed, we found that 'instructions to authors' pages provided limited and varied information for preparing a case report. There is a need for consensus, and more consistent guidance for authors of case report.     

High and Low Stimulus-Driven Conflict Engage Segregated Brain Networks,Not Quantitatively Different Resources

Task-irrelevant information is constantly present in our environment and may interfere with the processing of the information necessary to achieve goal-directed behavior. While task goals determine which information must be suppressed, the demand for inhibitory control depends on the strength of the interference induced by incoming, task-irrelevant information. Whether the same or distinct inhibitory processes are engaged to suppress various degrees of interference from task-irrelevant information remains largely unresolved. We investigated this question by manipulating the strength of the conflict induced by automatic word reading in a classical color Stroop task. High conflict was induced by presenting words in participant’s native language and low conflict by presenting words in a less familiar language. Behavioral performance and electrical neuroimaging analyses of event-related potentials to the words were analyzed following a two-by-two within-subject design with factors conflict strength (high; low) and color word/word ink congruency (congruent; incongruent). Behaviorally, we observed a significant conflict strength × congruency driven by a smaller Stroop effect in the low- than high conflict condition. Electrophysiologically, we observed a significant conflict strength × congruency interaction at the topographic level during the period of the N450 components, indicative of the engagement of distinct configurations of brain networks. No such interaction was found at the level of response strength. Electrical sources analyses localized the topographic effect within the anterior cingulate cortex and basal ganglia, left middle frontal and occipital areas. We interpret our results in terms of qualitatively distinct executive mechanisms for reactive inhibitory control in conditions of high versus low stimulus-driven conflict.     

Clinical Features,Non-Infectious Manifestations and Survival Analysis of 161 Children with Primary Immunodeficiency in Mexico: A Single Center Experience Over two Decades

Purpose

The hallmark of Primary immunodeficiencies (PID) is unusual infection, although other immunological non-infectious manifestations such as autoimmunity, allergy and cancer are often present. Most published reports focus on one disease or defect groups, so that a global prevalence of non-infectious manifestations of PID is hard to find. We aimed to describe the clinical features of our pediatric patients with PID, as well as the frequency and evolution of allergy, cancer and autoimmunity.

Methods

We reviewed all the available charts of patients being followed for PID from 1991 to the spring of 2012 at the National Institute of Pediatrics, Mexico City, to describe their demographic, clinical and laboratory features. Their diagnoses were established by pediatric immunologists in accordance to ESID criteria, including routine immunological workup and specialized diagnostic assays. We divided patients by decade of diagnosis to analyze their survival curves.

Results

There were 168 charts available, from which we excluded one duplicate and six equivocal diagnoses. We studied the charts of 161 PID patients (68 % male, 86 % alive), mostly from the center of the country, with a positive family history in 27 % and known consanguinity in 11 %. Eighty percent of the patients were diagnosed during the last decade. Current median age was 124 months; median age at onset of infections, 12 months; median age at diagnosis, 52 months; median age at death, 67.5 months. Severe infection and bleeding were the cause of 22 deaths. Eighty-six percent of all patients had at least one infection, while non-infectious manifestations had a global prevalence of 36 %, namely: autoimmunity 19 %, allergies 17 %, and cancer 2.4 %. Survival curves were not significantly different when compared by decade of diagnosis.

Conclusions

Compared to other registry reports, we found a lower prevalence of antibody defects, and of associated allergy and cancer. We could only locate two isolated IgA deficiencies and four cases of cancer among our PID patients. Although antibody defects are the most prevalent group (30 %), the distribution we found is similar to that reported in Iran, Kuwait, Egypt and Taiwan, with a close 27 % share for phagocyte defects, and 26 % for the formerly called “well-defined” syndromes. Of note, autoimmune and inflammatory complications are high among our patients with chronic granulomatous disease, as has been reported in both the United States and Japan, but not in Europe.

     

The Effect of Foot Strike Pattern on Achilles Tendon Load During Running

In this study we compared Achilles tendon loading parameters during barefoot running among females with different foot strike patterns using open-source computer muscle modeling software to provide dynamic simulations of running. Muscle forces of the gastrocnemius and soleus were estimated from experimental data collected in a motion capture laboratory during barefoot running for 11 runners utilizing a rearfoot strike (RFS) and 8 runners utilizing a non-RFS (NRFS) pattern. Our results show that peak Achilles tendon force occurred earlier in stance phase (p = 0.007), which contributed to a 15% increase in average Achilles tendon loading rate among participants adopting a NRFS pattern (p = 0.06). Stance time, step length, and the estimated number of steps per mile were similar between groups. However, runners with a NRFS pattern experienced 11% greater Achilles tendon impulse each step (p = 0.05) and nearly significantly greater Achilles tendon impulse per mile run (p = 0.06). This difference equates to an additional 47.7 body weights for each mile run with a NRFS pattern. Runners considering a NRFS pattern may want to account for these novel stressors and adapt training programs accordingly.     

Multisensory integration of drumming actions: musical expertise affects perceived audiovisual asynchrony

We investigated the effect of musical expertise on sensitivity to asynchrony for drumming point-light displays, which varied in their physical characteristics (Experiment 1) or in their degree of audiovisual congruency (Experiment 2). In Experiment 1, 21 repetitions of three tempos × three accents × nine audiovisual delays were presented to four jazz drummers and four novices. In Experiment 2, ten repetitions of two audiovisual incongruency conditions × nine audiovisual delays were presented to 13 drummers and 13 novices. Participants gave forced-choice judgments of audiovisual synchrony. The results of Experiment 1 show an enhancement in experts’ ability to detect asynchrony, especially for slower drumming tempos. In Experiment 2 an increase in sensitivity to asynchrony was found for incongruent stimuli; this increase, however, is attributable only to the novice group. Altogether the results indicated that through musical practice we learn to ignore variations in stimulus characteristics that otherwise would affect our multisensory integration processes.     

Molecular detection and genetic characterization of kobuviruses and astroviruses in asymptomatic local pigs in East Africa

In this study, swine fecal specimens (n = 251) collected from nursing and weaned piglets raised under smallholder production systems were screened for the presence of kobuviruses by RT-PCR. Porcine kobuviruses were detected in 13.1 % (33/251) of the samples. We demonstrated that porcine kobuvirus infections exist in indigenous pigs in Kenya and Uganda and that the prevalence was higher in young piglets than older pigs: nursing piglets (15 %), post-weaning (3-month-old) pigs (17 %), 4-month-old pigs (10 %). Genetic analysis of the partial RNA-dependent RNA polymerase (RdRp) region (690 nt) revealed that kobuviruses circulating in East Africa are diverse, sharing nucleotide sequence identities ranging from 89.7 to 99.1 % and 88 to 92.3 % among them and with known porcine kobuviruses, respectively. The nucleotide sequence identities between our kobuvirus strains and those of human, bovine and canine kobuviruses were 69.4-70.7 %, 73.1-74.4 % and 67-70.7 %, respectively. Additionally, upon sequencing selected samples that showed consistent 720-bp RT-PCR bands while using the same primer set, we detected porcine astroviruses in our samples belonging to type 2 and type 3 mamastroviruses. To our knowledge, this study reports the first detection and molecular analysis of both porcine kobuviruses and astroviruses in an African region. Further studies are required to determine the role of these viruses in gastrointestinal infections of pigs in this region and to determine the genetic diversity of the circulating strains to develop accurate diagnostic tools and implement appropriate control strategies.     

Cone beam computed tomography of plastinated hearts for instruction of radiological anatomy

Radiological anatomy education is an important aspect of the medical curriculum. The purpose of this study was to establish and demonstrate the use of plastinated anatomical specimens, specifically human hearts, for use in radiological anatomy education. Four human hearts were processed with routine plastination procedures at room temperature. Specimens were subjected to cone beam computed tomography and a graphics program (ER3D) was applied to generate 3D cardiac models. A comparison was conducted between plastinated hearts and their corresponding computer models based on a list of morphological cardiac features commonly studied in the gross anatomy laboratory. Results showed significant correspondence between plastinations and CBCT-generated 3D models (98 %; p < .01) for external structures and 100 % for internal cardiac features, while 85 % correspondence was achieved between plastinations and 2D CBCT slices. Complete correspondence (100 %) was achieved between key observations on the plastinations and internal radiological findings typically required of medical student. All pathologic features seen on the plastinated hearts were also visualized internally with the CBCT-generated models and 2D slices. These results suggest that CBCT-derived slices and models can be successfully generated from plastinated material and provide accurate representations for radiological anatomy education.     

Greek language processing in naive and skilled readers: Functional properties of the VWFA investigated with ERPs

Neuroimaging studies have provided evidence that the orthographic properties of linguistic stimuli are processed within the visual word form area (VWFA) localised in the left inferotemporal cortex (Cohen & Dehaene, 2004). It is not, however, clear in the literature whether this region responds preferentially to words, distinguishing them from pseudowords, or whether the pseudowords are distinguished from letter-strings on the basis of their orthographic regularity, or again, whether the VWFA distinguishes letters from numbers or from visual stimuli such as chequerboards. Very recently, it has been claimed that there is no evidence that the ill-named VWFA changes its responsiveness during or after reading acquisition (Price & Devlin, 2004). In order to simulate a condition of pre-reading ability in adult readers, we performed this study, in which we compared processing of Greek words and legal pseudowords in mother-tongue Greeks (skilled readers) and monolingual Italian individuals (naive readers) who had no familiarity with the Greek alphabet. ERPs were recorded while volunteers were engaged in a task involving the identification and response to target letters contained within Greek words or pseudowords. The response speed was identical in the two groups (550 vs. 557 ms). ERP data showed that at 165 ms post-stimulus (N1 component) the left lateral-occipital scalp area, probably corresponding to the left ventral occipitotemporal sulcus, discriminates letters of a familiar alphabet, while an unknown alphabet also activates the homologous right-hemispheric region more than 100 ms later. This suggests that the VWFA discriminates nonalphabetic symbols from letter-strings. An analysis of the N2 component showed an increase in the activation of the same region at about 285 ms post-stimulus during the processing of words rather than pseudowords in skilled readers, thus supporting the view that the VWFA discriminates words on the basis of their familiarity. The data seem to suggest that the VWFA is alphabet-specific and that it is based on the shaping of visual area activity during acquisition of the ability to read a given symbolic code.     

QRSE: a Novel Metric for the Evaluation of Trainee Radiologist Reporting Skills

Diagnostic radiology training programs must produce highly skilled diagnostic radiologists capable of interpreting radiological examinations and communicating results to clinicians. Established training performance tools evaluate interpretive skills, but trainees' competency in reporting skills is also essential. Our semi-automated passive electronic tool entitled the Quantitative Reporting Skills Evaluation (QRSE) allows radiology training programs to evaluate the quantity of edits made to trainee preliminary reports by attending physicians as a metric to evaluate trainee reporting performance. Consecutive report pairs and metadata extracted from the radiology information system were anonymized and exported to a MySQL database. To perform the QRSE, for each report pair, open source software was first utilized to calculate the Levenshtein Percent (LP), the percent of character changes required to convert each preliminary report to its corresponding final report. The average LP (ALP), ALP for each trainee, and standard deviations were calculated. Eighty-four trainees and 56 attending radiologists interpreted 228,543 radiological examinations during the study period. The overall ALP was 6.38 %. Trainee-specific ALPs ranged from 1.1 to 15.3 %. Among trainee-specific ALPs, the standard deviation was 3.7 %. Our analysis identified five trainees with trainee-specific ALPs above 2 standard deviations from the mean and 14 trainees with trainee-specific ALPs less than 1 standard deviation below the mean. The QRSE methodology allows for the passive, quantitative, and longitudinal evaluation of the reporting skills of trainees during diagnostic radiology residency training. The QRSE identifies trainees with high and low levels of edits to their preliminary reports, as a marker for trainee overall reporting skills, and thus represents a novel performance metric for radiology training programs.     

Three-dimensional segmentation of retroperitoneal masses using continuous convex relaxation and accumulated gradient distance for radiotherapy planning

An innovative algorithm has been developed for the segmentation of retroperitoneal tumors in 3D radiological images. This algorithm makes it possible for radiation oncologists and surgeons semiautomatically to select tumors for possible future radiation treatment and surgery. It is based on continuous convex relaxation methodology, the main novelty being the introduction of accumulated gradient distance, with intensity and gradient information being incorporated into the segmentation process. The algorithm was used to segment 26 CT image volumes. The results were compared with manual contouring of the same tumors. The proposed algorithm achieved 90 % sensitivity, 100 % specificity and 84 % positive predictive value, obtaining a mean distance to the closest point of 3.20 pixels. The algorithm’s dependence on the initial manual contour was also analyzed, with results showing that the algorithm substantially reduced the variability of the manual segmentation carried out by different specialists. The algorithm was also compared with four benchmark algorithms (thresholding, edge-based level-set, region-based level-set and continuous max-flow with two labels). To the best of our knowledge, this is the first time the segmentation of retroperitoneal tumors for radiotherapy planning has been addressed.     

Effect of root bark extract of Berberis vulgaris L. on Leishmania major on BALB/c mice

Leishmaniasis is one of the most important diseases transmitted by arthropod. This disease is present in 88 countries. Approximately 400 million people are at risk, and 12 million are involved. We aimed to examine the application of ethanolic extract of the root bark of Berberis vulgaris L. for treatment of mice infected with cutaneous leishmaniasis. At first, 40 BALB/c mice were infected to Leishmania major promastigotes and were divided in two groups A and B. Then, each of A and B groups were divided to two subgroups. Mice from subgroup A1 were treated with 10 % root bark alcoholic extract, and mice from subgroup A2 were treated with only alcohol (control). Mice from subgroup B1 were treated with 20 % root bark alcoholic extract, and mice from subgroup B2 were treated with only alcohol (control). The 90 % recovery was found in the mice treated with 20 % root bark extract, and 55 % recovery was found with 10 % root bark extract, but in the control group, 0 % recovery was found. The results of our study showed that the lotion of root bark extract has good suppression effects on parasites. Therefore, it might be a pro for developing new antileishmanial drugs.     

Anatomical variations of the celiac trunk: cadaveric and radiological study

Purpose

The celiac trunk (CT) is a vascular structure of the upper abdomen which gives off the left gastric artery (LGA), the splenic artery and the common hepatic artery. This study aims to compare the vascular patterns of the CT of two different samples (cadaveric and radiological) and to propose a simple classification of CT variations based on previous studies and our results.

Materials and methods

To perform this study we examined 43 adult cadavers, 24 males and 19 females, ages ranged from 69 to 92. In addition, we analysed 596 MDCT (multidetector computed tomographic) angiography examinations of 430 males and 166 females, ages ranged from 42 to 82.

Results

According to the classification proposed, results were divided into Type I or complete CT (578/639 cases, 90.5 %), Type II or incomplete CT (61/639 cases, 9.5 %), Type III or absence of CT and Type IV or celiacomesenteric trunk with no cases reported. Type I was divided into Type Ia or bifurcated trunk with LGA arising first (368/639 cases, 57.6 %), Type Ib or trifurcated trunk (205/639 cases, 32.1 %) and Type Ic or tetrafurcated trunk with an extra branch (5/639 cases, 0.8 %). Type II included hepatosplenic (29/639 cases, 4.5 %), gastroplenic (32/639, 5 %) and hepatogastric trunks (0/639, 0 %) which represented Types IIa, IIb and IIc respectively.

Conclusions

No significant differences were found between the cadaveric and radiological samples. Gender did not appear to be related to any variability of the structures either. A new, simple and complete classification of the anatomical variations of the CT is proposed.

     

The Ensembl Web site: mechanics of a genome browser

The Ensembl Web site (http://www.ensembl.org/) is the principal user interface to the data of the Ensembl project, and currently serves >500,000 pages (approximately 2.5 million hits) per week, providing access to >80 GB (gigabyte) of data to users in more than 80 countries. Built atop an open-source platform comprising Apache/mod_perl and the MySQL relational database management system, it is modular, extensible, and freely available. It is being actively reused and extended in several different projects, and has been downloaded and installed in companies and academic institutions worldwide. Here, we describe some of the technical features of the site, with particular reference to its dynamic configuration that enables it to handle disparate data from multiple species.     

Combining risk communication strategies to simultaneously convey the risks of four diseases associated with physical inactivity to socio-demographically diverse populations

A single risk factor can increase the risk of developing multiple diseases, but most risk communication research has been conducted in the context of a single disease. We explored which combination of three recommended risk communication strategies is most effective in simultaneously conveying risk estimates of four diseases associated with physical inactivity: colon cancer, stroke, diabetes, and heart disease. Participants (N = 1161, 50% no college experience, 50% racial/ethnic minority) were shown hypothetical risk estimates for each of the four diseases. All four diseases were placed at varying heights on 1 of 12 vertical bar charts (i.e., “risk ladders”) to indicate their respective probabilities. The risk ladders varied in a 2 (risk reduction information: present/absent) × 2 (numerical format: words/words and numbers) × 3 (social comparison information: none/somewhat higher than average/much higher than average) full factorial design. Participants were randomly assigned to view one of the risk ladders and then completed a questionnaire assessing message comprehension, message acceptance, physical activity-related risk and efficacy beliefs, and physical activity intentions. Higher message acceptance was found among (1) people who received risk reduction information versus those who did not (p = .01), and (2) people who did not receive social comparison information versus those told that they were at higher than average risk (p = .03). Further, absolute cognitive perceived risk of developing “any of the diseases shown in the picture” was higher among people who did not receive social comparison information (p = .03). No other main effects and only very few interactions with demographic variables were found. Combining recommended risk communication strategies did not improve or impair key cognitive or affective precursors of health behavior change. It might not be necessary to provide people with extensive information when communicating risk estimates of multiple diseases.     

Give Me a Sign: Decoding Complex Coordinated Hand Movements Using High-Field fMRI

Decoding movements from the human cortex has been a topic of great interest for controlling an artificial limb in non-human primates and severely paralyzed people. Here we investigate feasibility of decoding gestures from the sensorimotor cortex in humans, using 7 T fMRI. Twelve healthy volunteers performed four hand gestures from the American Sign Language Alphabet. These gestures were performed in a rapid event related design used to establish the classifier and a slow event-related design, used to test the classifier. Single trial patterns were classified using a pattern-correlation classifier. The four hand gestures could be classified with an average accuracy of 63 % (range 35–95 %), which was significantly above chance (25 %). The hand region was, as expected, the most active region, and the optimal volume for classification was on average about 200 voxels, although this varied considerably across individuals. Importantly, classification accuracy correlated significantly with consistency of gesture execution. The results of our study demonstrate that decoding gestures from the hand region of the sensorimotor cortex using 7 T fMRI can reach very high accuracy, provided that gestures are executed in a consistent manner. Our results further indicate that the neuronal representation of hand gestures is robust and highly reproducible. Given that the most active foci were located in the hand region, and that 7 T fMRI has been shown to agree with electrocorticography, our results suggest that this confined region could serve to decode sign language gestures for intracranial brain–computer interfacing using surface grids.     

Limited role for histopathological examination of re-excision specimens of completely excised melanomas

The Dutch melanoma guideline advises to examine one central block of the re-excision scar in case of a complete primary excision. To increase the evidence for this recommendation, we re-evaluated how often residual melanoma was found in re-excision specimens of a large series of completely excised melanomas. Of 1,209 Dutch melanoma cases, pathology reports of primary excisions were reviewed. Presence of melanoma in the margins was scored. All melanomas with a complete primary excision were included and pathology reports of re-excisions were reviewed. Presence of residual melanoma in the re-excision specimen and the number of blocks were scored. Slides of re-excision specimens containing residual melanoma were reviewed. Eventually, in four out of 812 melanomas (0.5 %) with a complete primary excision, residual melanoma was found in the re-excision specimen. The free margins of the primary melanomas in these cases ranged from 0.5–3.5 mm. In one case, the margin for melanoma in situ was 0.2 mm. In <1 % of initially completely excised melanomas, residual melanoma was found in the re-excision specimen. Histopathological examination of these re-excision specimens may not be cost-efficient. Our findings even imply that a re-excision could safely be omitted in selected cases of completely excised melanomas.     

A Visual Query-by-Example Image Database for Chest CT Images: Potential Role as a Decision and Educational Support Tool for Radiologists

Primary reading or further evaluation of diagnostic imaging examination often needs a comparison between the actual findings and the relevant prior images of the same patient or similar radiological data found in other patients. This support is of clinical importance and may have significant effects on physicians examination reading efficiency, service-quality, and work satisfaction. We developed a visual query-by-example image database for storing and retrieving chest CT images by means of a visual browser Image Management Environment (IME) and tested its retrieval efficiency. The visual browser IME included four fundamental features (segmentation, indexing, quick load and recall, user-friendly interface) in an integrated graphical environment for a user-friendly image database management. The system was tested on a database of 2000 chest CT images, randomly chosen from the digital archives of our institutions. A sample of eight heterogeneous images were used as queries and, for each of them a team of three expert radiologists selected the most similar images from the database (a set of 15 images containing similar abnormalities in the same position of the query). The sensitivity and the positive predictive factor, both averaged over the 8 test queries and 15 answers, were respectively 0.975 and 0.91 The IME system is currently under evaluation at our institutions as an experimental application. We consider it a useful work-in-progress tool for clinical practice facilitating searches for a variety of radiological tasks.