Search

Background Large language models have already demonstrated potential in medical text processing. GPT-4V, a large vision-language model from OpenAI, has shown potential for medical imaging, yet a quantitative analysis is lacking. Purpose To quantitatively assess the performance of GPT-4V in interpreting radiologic images using unseen data. Materials and Methods This retrospective study included single representative abnormal and healthy control images from neuroradiology, cardiothoracic radiology, and musculoskeletal radiology (CT, MRI, radiography) to generate reports using GPT-4V via the application programming interface from February to March 2024. The factual correctness of free-text reports and the performance in detecting abnormalities in binary classification tasks were assessed using accuracy, sensitivity, and specificity. The binary classification performance was compared with that of a first-year nonradiologist in training and four board-certified radiologists. Results A total of 515 images in 470 patients (median age, 61 years [IQR, 44-71 years]; 267 male) were included, of which 345 images were abnormal. GPT-4V correctly identified the imaging modality and anatomic region in 100% (515 of 515) and 99.2% (511 of 515) of images, respectively. Diagnostic accuracy in free-text reports was between 0% (0 of 33 images) for pneumothorax (CT and radiography) and 90% (45 of 50 images) for brain tumor (MRI). In binary classification tasks, GPT-4V showed sensitivities between 56% (14 of 25 images) for ischemic stroke and 100% (25 of 25 images) for brain hemorrhage and specificities between 8% (two of 25 images) for brain hemorrhage and 52% (13 of 25 images) for pneumothorax, compared with a pooled sensitivity of 97.2% (1103 of 1135 images) and pooled specificity of 97.2% (1084 of 1115 images) for the human readers across all tasks. The model exhibited a clear tendency to overdiagnose abnormalities, with 86.5% (147 of 170 images) and 67.7% (151 of 223 images) false-positive rates for the free-text and binary classification tasks, respectively. Conclusion GPT-4V, in its earliest version, recognized medical image content and reliably determined the modality and anatomic region from single images. However, GPT-4V failed to detect, classify, or rule out abnormalities in image interpretation. © RSNA, 2024 Supplemental material is available for this article.

Purpose: To evaluate prospectively the accuracy of indirect magnetic resonance (MR) arthrography for supraspinatus tendon tears during neutral positioning or abduction and external rotation (ABER) and neutral positioning., Materials and Methods: Informed consent was obtained in all patients, and the study was approved by the institutional review board. Indirect MR arthrography of the shoulder was performed in 51 symptomatic patients (14 female, 37 male; mean age, 47 years) in the neutral position (set 1) and in the neutral and ABER positions (set 2). Two readers independently interpreted both sets, and diagnoses were compared with arthroscopic findings. Diagnostic accuracy was calculated, and 95% confidence intervals were used to detect significant differences between sets. Diagnostic confidence was recorded by using a three-level confidence score. Differences between sets were evaluated by using the Wilcoxon signed rank test. Interobserver agreement was determined separately for each set and for all diagnoses, full-thickness tears, and partial-thickness tears., Results: For full-thickness tears, there was no benefit to reading set 2. For reader 1, sensitivity and specificity were 95% and 100%, respectively, for set 1 and 100% and 100%, respectively, for set 2. For reader 2, sensitivity and specificity were 80% and 100%, respectively, for set 1 and 100% and 100%, respectively, for set 2. For partial-thickness tears, sensitivity was significantly higher after reading set 2. For reader 1, sensitivity and specificity were 71% and 88%, respectively, for set 1 and 93% and 100%, respectively, for set 2. For reader 2, sensitivity and specificity were 50% and 88%, respectively, for set 1 and 86% and 94%, respectively, for set 2. For both readers, diagnostic confidence for partial-thickness tears was significantly higher after reading set 2. After the interpretation of set 2, kappa values increased from 0.35 to 1.00 for full-thickness tears and from 0.12 to 0.63 for partial-thickness tears., Conclusion: Indirect MR arthrography with supplementary images obtained with patients in the ABER position significantly improved sensitivity and increased diagnostic confidence for partial-thickness tears of the supraspinatus tendon. Interobserver agreement was improved for both full- and partial-thickness tears., (RSNA, 2006)

Purpose: To compare image quality and estimated dose for chest radiographs obtained by using a cesium iodide-amorphous silicon flat-panel detector at fixed tube voltage and detector entrance dose with and without additional 0.3-mm copper filtration., Materials and Methods: The study was approved by the institutional ethics committee. All prospectively enrolled patients signed the written consent form. Chest radiographs in two projections were acquired at 125-kVp tube voltage and 2.5-microGy detector entrance dose. The experimental group (38 patients) was imaged with 0.3-mm copper filtration; the control group (38 patients) was imaged without copper filtration. An additional 12 patients were imaged with and without copper filtration and served as paired subject-controls. Three readers blinded to group and clinical data independently evaluated the radiographs for image quality on a digital display system. Twelve variables (six for each radiographic projection) were assigned scores on a seven-point ordinal scale. Scores between experimental and control groups were compared: Logistic regression analysis and Mann-Whitney U test were used for unpaired patients; and Wilcoxon and McNemar test, for paired patients. In all, 72 comparisons were determined (36 [12 variables x three readers] for unpaired patients and 36 for paired patients). In a phantom study, radiation burden of experimental protocol was compared with that of control protocol by using Monte Carlo calculations., Results: For 70 of 72 comparisons, digital radiographs obtained with copper filtration were of similar image quality as radiographs obtained without copper filtration (P = .123 to P > .99). For two of 72 comparisons, one observer judged the experimental protocol superior to the control protocol (P = .043, P = .046). Patient dose reduction estimated with Monte Carlo calculations was 31%. Use of copper filtration increased exposure times by 48% for posteroanterior views and by 34% for lateral views., Conclusion: Subjectively equivalent chest radiographic image quality was found with estimated 30% dose reduction after addition of 0.3-mm copper filtration with flat-panel cesium iodide-amorphous silicon technology.

Purpose: To evaluate the safety and efficacy of gadoxetic acid disodium-enhanced magnetic resonance (MR) imaging for the detection of focal liver lesions, with results of histopathologic examination and/or intraoperative ultrasonography used as a standard of reference., Materials and Methods: One hundred sixty-nine patients who were known to have or suspected of having focal liver lesions and were scheduled for liver surgery were included in this study. Results in 131 patients could be included in the efficacy analysis. MR imaging was performed before and immediately and 20 minutes after bolus injection of 0.025 mmol/kg of the liver-specific hepatobiliary contrast agent gadoxetic acid. T1-weighted gradient-echo (with and without fat saturation and including dynamic data sets) and T2-weighted fast spin-echo/turbo spin-echo sequences were performed. All images were evaluated on site and by three independent and blinded off-site reviewers. Lesion matching based on the standard-of-reference results was performed. Differences in lesion detection with precontrast and with postcontrast MR images were assessed with the two-sided Wilcoxon signed rank test., Results: Gadoxetic acid was well tolerated. In the on-site review, the number of patients in whom all lesions were correctly matched increased from 89 of 129 patients at precontrast MR imaging to 103 of 129 patients at postcontrast MR imaging. In the off-site evaluation, the number of patients in whom all lesions were correctly matched and the corresponding sensitivity values increased from 72 (55.8%), 68 (52.7%), and 66 (51.2%) with the precontrast images to 88 (68.2%), 69 (53.5%), and 76 (58.9%) with the postcontrast images for readers 1, 2, and 3, respectively. Two of the three blinded readers showed a statistically significant difference in lesion detection between precontrast and postcontrast MR imaging (P <.001 and P =.008). A large number of additionally correctly detected and localized lesions were smaller than 1 cm., Conclusion: MR imaging with gadoxetic acid is safe and improves lesion detection and localization., (Copyright RSNA, 2004)

Purpose: Results with different doses of gadobenate dimeglumine and gadopentetate dimeglumine were compared at magnetic resonance (MR) angiography of the renal arteries. The signal-to-noise ratio (SNR) was evaluated as a quantitative measure of image quality., Materials and Methods: Sixty consecutive patients (age range, 24-81 years; mean age, 65 years) underwent intraarterial digital subtraction angiography (DSA) and contrast material-enhanced time-resolved MR angiography. DSA was the standard of reference. Fifteen patients received gadopentetate dimeglumine at doses of 0.2 or 0.1 mmol per kilogram of body weight. Fifteen patients received gadobenate dimeglumine at doses of 0.05 or 0.1 mmol/kg. The SNR was calculated in the aorta and both main renal arteries. The number and degree of stenoses of the renal arteries and accessory vessels were evaluated by four observers., Results: SNRs with gadobenate dimeglumine at a dose of 0.1 mmol/kg were significantly superior to those with gadopentetate dimeglumine at a dose of 0.1 mmol/kg. Differences were not statistically significant between the SNRs in the other groups. Eleven (85%) of 13 hemodynamically significant renal artery stenoses were detected correctly with MR angiography as were 22 (85%) of 26 accessory renal arteries., Conclusion: SNRs with gadobenate dimeglumine were higher than those with gadopentetate dimeglumine, but in most cases the differences in SNRs were not statistically significant.

Purpose: To assess the feasibility of magnetic resonance (MR) imaging-guided stent placement in iliac arterial stenoses., Materials and Methods: Thirteen patients with 14 iliac arterial stenoses were examined prospectively. Angioplasty was performed through a femoral sheath by using a conventional 1.5-T MR imaging system. Stents and catheters were visualized on the basis of their artifacts. Nitinol stents were placed with gradient-echo MR imaging guidance. Angioplasty balloons were inflated with gadolinium-based contrast material. Results were evaluated clinically and with both digital subtraction angiography (DSA) and contrast material-enhanced MR angiography., Results: Ten of 13 patients were treated with technical success by using MR imaging-guided intervention alone. Three patients were treated with additional fluoroscopic guidance, because complications (ie, panic attack, subintimal recanalization, and stent misplacement) occurred with MR guidance. The quality of the postinterventional contrast-enhanced MR angiograms of three of 12 lesions with stents was limited owing to stent-induced signal loss of the lumen. The mean stenosis degree after stent placement was significantly higher at contrast-enhanced MR angiography than at DSA (24.6% vs 6.2%). The mean MR imaging-guided procedure time was 74 minutes., Conclusion: MR imaging-guided stent placement in iliac arteries is feasible in select patients. The presented technique has limitations-that is, long procedure times, lack of real-time monitoring, and stent artifacts-that necessitate further modifications before it can be recommended for clinical use.

Purpose: To evaluate signal intensity changes influencing assessment of stent patency at contrast material-enhanced magnetic resonance (MR) angiography., Materials and Methods: By using an in vitro model, 14 stents-nine nitinol, one tantalum, two stainless steel, and two cobalt alloy-were investigated regarding their appearance at MR imaging. A vascular phantom consisting of tubes filled with 2.00 mmol/L gadopentetate dimeglumine in saline solution was studied in different orientations within the magnetic field. Imaging was performed with a fast three-dimensional gradient-echo sequence (4. 70/1.89 [repetition time msec/echo time msec]). Relative signal intensity reduction within the stents and the degree of artificial narrowing of the stent lumen were calculated., Results: The stent lumen was visible within 13 stents. A total signal void inside the stent lumen appeared in only one cobalt alloy stent. Artificial narrowing of the diameter was less than 33% in 10 of 14 stents. The tantalum stent and four nitinol stents seemed best suited for contrast-enhanced MR angiography. A bandlike artifact occurred at the ends of the stents when positioned along the readout direction., Conclusion: To differentiate between artifacts and stenoses, knowledge of the degree of signal intensity reduction and artificial lumen narrowing within vascular stents is essential. Stent geometry, relative orientation to the magnetic field, and alloy composition influence signal intensity alteration within the stent lumen.

Purpose: To assess the diagnostic performance of an active-matrix flat-panel x-ray detector for reduced-dose imaging of simulated arthritic lesions., Materials and Methods: A digital x-ray detector based on cesium iodide and amorphous silicon technology with a panel size of 43 x 43 cm, matrix of 3,000 x 3,000 pixels, pixel size of 143 micrometer, and digital output of 14 bits was used. State-of-the-art screen-film radiographs were compared with digital images obtained at doses equivalent to those obtained with system speeds of 400, 560, and 800. The phantom was composed of a human hand skeleton on an acrylic plate with drilled holes simulating bone erosions of different diameters and depths. Results of four independent observers were evaluated with receiver operating characteristic curve analysis., Results: The cesium iodide and amorphous silicon detector resulted in better diagnostic performance than did the screen-film combination, with the dose being the same for both modalities (P <.05). For digital images obtained at reduced doses, no significant differences were found., Conclusion: The improved diagnostic performance with digital radiographs obtained with the cesium iodide and amorphous silicon detector suggests that this detector technology holds promise in terms of dose reduction for specific diagnostic tasks, without loss of diagnostic accuracy.