RESUMEN
Segmentation of infarcts is clinically important in ischemic stroke management and prognostication. It is unclear what role the combination of DWI, ADC, and FLAIR MRI sequences provide for deep learning in infarct segmentation. Recent technologies in model self-configuration have promised greater performance and generalizability through automated optimization. We assessed the utility of DWI, ADC, and FLAIR sequences on ischemic stroke segmentation, compared self-configuring nnU-Net models to conventional U-Net models without manual optimization, and evaluated the generalizability of results on an external clinical dataset. 3D self-configuring nnU-Net models and standard 3D U-Net models with MONAI were trained on 200 infarcts using DWI, ADC, and FLAIR sequences separately and in all combinations. Segmentation results were compared between models using paired t-test comparison on a hold-out test set of 50 cases. The highest performing model was externally validated on a clinical dataset of 50 MRIs. nnU-Net with DWI sequences attained a Dice score of 0.810 ± 0.155. There was no statistically significant difference when DWI sequences were supplemented with ADC and FLAIR images (Dice score of 0.813 ± 0.150; p = 0.15). nnU-Net models significantly outperformed standard U-Net models for all sequence combinations (p < 0.001). On the external dataset, Dice scores measured 0.704 ± 0.199 for positive cases with false positives with intracranial hemorrhage. Highly optimized neural networks such as nnU-Net provide excellent stroke segmentation even when only provided DWI images, without significant improvement from other sequences. This differs from-and significantly outperforms-standard U-Net architectures. Results translated well to the external clinical environment and provide the groundwork for optimized acute stroke segmentation on MRI.
RESUMEN
Background Large language models (LLMs) for medical applications use unknown amounts of energy, which contribute to the overall carbon footprint of the health care system. Purpose To investigate the tradeoffs between accuracy and energy use when using different LLM types and sizes for medical applications. Materials and Methods This retrospective study evaluated five different billion (B)-parameter sizes of two open-source LLMs (Meta's Llama 2, a general-purpose model, and LMSYS Org's Vicuna 1.5, a specialized fine-tuned model) using chest radiograph reports from the National Library of Medicine's Indiana University Chest X-ray Collection. Reports with missing demographic information and missing or blank files were excluded. Models were run on local compute clusters with visual computing graphic processing units. A single-task prompt explained clinical terminology and instructed each model to confirm the presence or absence of each of the 13 CheXpert disease labels. Energy use (in kilowatt-hours) was measured using an open-source tool. Accuracy was assessed with 13 CheXpert reference standard labels for diagnostic findings on chest radiographs, where overall accuracy was the mean of individual accuracies of all 13 labels. Efficiency ratios (accuracy per kilowatt-hour) were calculated for each model type and size. Results A total of 3665 chest radiograph reports were evaluated. The Vicuna 1.5 7B and 13B models had higher efficiency ratios (737.28 and 331.40, respectively) and higher overall labeling accuracy (93.83% [3438.69 of 3665 reports] and 93.65% [3432.38 of 3665 reports], respectively) than that of the Llama 2 models (7B: efficiency ratio of 13.39, accuracy of 7.91% [289.76 of 3665 reports]; 13B: efficiency ratio of 40.90, accuracy of 74.08% [2715.15 of 3665 reports]; 70B: efficiency ratio of 22.30, accuracy of 92.70% [3397.38 of 3665 reports]). Vicuna 1.5 7B had the highest efficiency ratio (737.28 vs 13.39 for Llama 2 7B). The larger Llama 2 70B model used more than seven times the energy of its 7B counterpart (4.16 kWh vs 0.59 kWh) with low overall accuracy, resulting in an efficiency ratio of only 22.30. Conclusion Smaller fine-tuned LLMs were more sustainable than larger general-purpose LLMs, using less energy without compromising accuracy, highlighting the importance of LLM selection for medical applications. © RSNA, 2024 Supplemental material is available for this article.
Asunto(s)
Radiografía Torácica , Estudios Retrospectivos , Humanos , Radiografía Torácica/métodosRESUMEN
PURPOSE: To evaluate whether a commercial AI tool for intracranial hemorrhage (ICH) detection on head CT exhibited sociodemographic biases. METHODS: Our retrospective study reviewed 9736 consecutive, adult non-contrast head CT scans performed between November 2021 and February 2022 in a single healthcare system. Each CT scan was evaluated by a commercial ICH AI tool and a board-certified neuroradiologist; ground truth was defined as final radiologist determination of ICH presence/absence. After evaluating the AI tool's aggregate diagnostic performance, sub-analyses based on sociodemographic groups (age, sex, race, ethnicity, insurance status, and Area of Deprivation Index [ADI] scores) assessed for biases. χ2 test or Fisher's exact tests evaluated for statistical significance with p ≤ 0.05. RESULTS: Our patient population was 50% female (mean age 60 ± 19 years). The AI tool had an aggregate accuracy of 93% [9060/9736], sensitivity of 85% [1140/1338], specificity of 94% [7920/ 8398], positive predictive value (PPV) of 71% [1140/1618] and negative predictive value (NPV) of 98% [7920/8118]. Sociodemographic biases were identified, including lower PPV for patients who were females (67.3% [62,441/656] vs. 72.7% [699/962], p = 0.02), Black (66.7% [454/681] vs. 73.2% [686/937], p = 0.005), non-Hispanic/non-Latino (69.7% [1038/1490] vs. 95.4% [417/437]), p = 0.009), and who had Medicaid/Medicare (69.9% [754/1078]) or Private (66.5% [228/343]) primary insurance (p = 0.003). Lower sensitivity was seen for patients in the third quartile of national (78.8% [241/306], p = 0.001) and state ADI scores (79.0% [22/287], p = 0.001). CONCLUSIONS: In our healthcare system, a commercial AI tool had lower performance for ICH detection than previously reported and demonstrated several sociodemographic biases.
RESUMEN
BACKGROUND: Radiographic diagnosis of necrotizing enterocolitis (NEC) is challenging. Deep learning models may improve accuracy by recognizing subtle imaging patterns. We hypothesized it would perform with comparable accuracy to that of senior surgical residents. METHODS: This cohort study compiled 494 anteroposterior neonatal abdominal radiographs (214 images NEC, 280 other) and randomly divided them into training, validation, and test sets. Transfer learning was utilized to fine-tune a ResNet-50 deep convolutional neural network (DCNN) pre-trained on ImageNet. Gradient-weighted Class Activation Mapping (Grad-CAM) heatmaps visualized image regions of greatest relevance to the pretrained neural network. Senior surgery residents at a single institution examined the test set. Resident and DCNN ability to identify pneumatosis on radiographic images were measured via area under the receiver operating curves (AUROC) and compared using DeLong's method. RESULTS: The pretrained neural network achieved AUROC of 0.918 (95% CI, 0.837-0.978) with an accuracy of 87.8% with five false negative and one false positive prediction. Heatmaps confirmed appropriate image region emphasis by the pretrained neural network. Senior surgical residents had a median area under the receiver operating curve of 0.896, ranging from 0.778 (95% CI 0.615-0.941) to 0.991 (95% CI 0.971-0.999) with zero to five false negatives and one to eleven false positive predictions. The deep convolutional neural network performed comparably to each surgical resident's performance (p > 0.05 for all comparisons). CONCLUSIONS: A deep convolutional neural network trained to recognize pneumatosis can quickly and accurately assist clinicians in promptly identifying NEC in clinical practice. LEVEL OF EVIDENCE: III (study type: Study of Diagnostic Test, study of nonconsecutive patients without a universally applied "gold standard").
RESUMEN
As the adoption of artificial intelligence (AI) systems in radiology grows, the increase in demand for greater bandwidth and computational resources can lead to greater infrastructural costs for healthcare providers and AI vendors. To that end, we developed ISLE, an intelligent streaming framework to address inefficiencies in current imaging infrastructures. Our framework draws inspiration from video-on-demand platforms to intelligently stream medical images to AI vendors at an optimal resolution for inference from a single high-resolution copy using progressive encoding. We hypothesize that ISLE can dramatically reduce the bandwidth and computational requirements for AI inference, while increasing throughput (i.e., the number of scans processed by the AI system per second). We evaluate our framework by streaming chest X-rays for classification and abdomen CT scans for liver and spleen segmentation and comparing them with the original versions of each dataset. For classification, our results show that ISLE reduced data transmission and decoding time by at least 92% and 88%, respectively, while increasing throughput by more than 3.72 × . For both segmentation tasks, ISLE reduced data transmission and decoding time by at least 82% and 88%, respectively, while increasing throughput by more than 2.9 × . In all three tasks, the ISLE streamed data had no impact on the AI system's diagnostic performance (all P > 0.05). Therefore, our results indicate that our framework can address inefficiencies in current imaging infrastructures by improving data and computational efficiency of AI deployments in the clinical environment without impacting clinical decision-making using AI systems.
RESUMEN
The large language model GPT-4 showed limited utility in generating BI-RADS assessment categories for factitious breast imaging reports containing findings and impression sections, with frequent incorrect BI-RADS category assignments and poor reproducibility in assigned BI-RADS categories across independent tests for the same report using the same prompt.
RESUMEN
Saliency maps are popularly used to "explain" decisions made by modern machine learning models, including deep convolutional neural networks (DCNNs). While the resulting heatmaps purportedly indicate important image features, their "trustworthiness," i.e., utility and robustness, has not been evaluated for musculoskeletal imaging. The purpose of this study was to systematically evaluate the trustworthiness of saliency maps used in disease diagnosis on upper extremity X-ray images. The underlying DCNNs were trained using the Stanford MURA dataset. We studied four trustworthiness criteria-(1) localization accuracy of abnormalities, (2) repeatability, (3) reproducibility, and (4) sensitivity to underlying DCNN weights-across six different gradient-based saliency methods (Grad-CAM (GCAM), gradient explanation (GRAD), integrated gradients (IG), Smoothgrad (SG), smooth IG (SIG), and XRAI). Ground-truth was defined by the consensus of three fellowship-trained musculoskeletal radiologists who each placed bounding boxes around abnormalities on a holdout saliency test set. Compared to radiologists, all saliency methods showed inferior localization (AUPRCs: 0.438 (SG)-0.590 (XRAI); average radiologist AUPRC: 0.816), repeatability (IoUs: 0.427 (SG)-0.551 (IG); average radiologist IOU: 0.613), and reproducibility (IoUs: 0.250 (SG)-0.502 (XRAI); average radiologist IOU: 0.613) on abnormalities such as fractures, orthopedic hardware insertions, and arthritis. Five methods (GCAM, GRAD, IG, SG, XRAI) passed the sensitivity test. Ultimately, no saliency method met all four trustworthiness criteria; therefore, we recommend caution and rigorous evaluation of saliency maps prior to their clinical use.
RESUMEN
Large language models (LLMs) hold immense potential to revolutionize radiology. However, their integration into practice requires careful consideration. Artificial intelligence (AI) chatbots and general-purpose LLMs have potential pitfalls related to privacy, transparency, and accuracy, limiting their current clinical readiness. Thus, LLM-based tools must be optimized for radiology practice to overcome these limitations. While research and validation for radiology applications remain in their infancy, commercial products incorporating LLMs are becoming available alongside promises of transforming practice. To help radiologists navigate this landscape, this AJR Expert Panel Narrative Review provides a multidimensional perspective on LLMs, encompassing considerations from bench (development and optimization) to bedside (use in practice). At present, LLMs are not autonomous entities that can replace expert decision-making, and radiologists remain responsible for the content of their reports. Patient-facing tools, particularly medical AI chatbots, require additional guardrails to ensure safety and prevent misuse. Still, if responsibly implemented, LLMs are well-positioned to transform efficiency and quality in radiology. Radiologists must be well-informed and proactively involved in guiding the implementation of LLMs in practice to mitigate risks and maximize benefits to patient care.
RESUMEN
Purpose To evaluate the robustness of an award-winning bone age deep learning (DL) model to extensive variations in image appearance. Materials and Methods In December 2021, the DL bone age model that won the 2017 RSNA Pediatric Bone Age Challenge was retrospectively evaluated using the RSNA validation set (1425 pediatric hand radiographs; internal test set in this study) and the Digital Hand Atlas (DHA) (1202 pediatric hand radiographs; external test set). Each test image underwent seven types of transformations (rotations, flips, brightness, contrast, inversion, laterality marker, and resolution) to represent a range of image appearances, many of which simulate real-world variations. Computational "stress tests" were performed by comparing the model's predictions on baseline and transformed images. Mean absolute differences (MADs) of predicted bone ages compared with radiologist-determined ground truth on baseline versus transformed images were compared using Wilcoxon signed rank tests. The proportion of clinically significant errors (CSEs) was compared using McNemar tests. Results There was no evidence of a difference in MAD of the model on the two baseline test sets (RSNA = 6.8 months, DHA = 6.9 months; P = .05), indicating good model generalization to external data. Except for the RSNA dataset images with an appended radiologic laterality marker (P = .86), there were significant differences in MAD for both the DHA and RSNA datasets among other transformation groups (rotations, flips, brightness, contrast, inversion, and resolution). There were significant differences in proportion of CSEs for 57% of the image transformations (19 of 33) performed on the DHA dataset. Conclusion Although an award-winning pediatric bone age DL model generalized well to curated external images, it had inconsistent predictions on images that had undergone simple transformations reflective of several real-world variations in image appearance. Keywords: Pediatrics, Hand, Convolutional Neural Network, Radiography Supplemental material is available for this article. © RSNA, 2024 See also commentary by Faghani and Erickson in this issue.
Asunto(s)
Determinación de la Edad por el Esqueleto , Aprendizaje Profundo , Niño , Humanos , Algoritmos , Redes Neurales de la Computación , Radiografía , Estudios Retrospectivos , Determinación de la Edad por el Esqueleto/métodosRESUMEN
As applications of AI in medicine continue to expand, there is an increasing focus on integration into clinical practice. An underappreciated aspect of this clinical translation is where the AI fits into the clinical workflow, and in turn, the outputs generated by the AI to facilitate clinician interaction in this workflow. For instance, in the canonical use case of AI for medical image interpretation, the AI could prioritize cases before clinician review or even autonomously interpret the images without clinician review. A related aspect is explainability - does the AI generate outputs to help explain its predictions to clinicians? While many clinical AI workflows and explainability techniques have been proposed, a summative assessment of the current scope in clinical practice is lacking. Here, we evaluate the current state of FDA-cleared AI devices for medical image interpretation assistance in terms of intended clinical use, outputs generated, and types of explainability offered. We create a curated database focused on these aspects of the clinician-AI interface, where we find a high frequency of "triage" devices, notable variability in output characteristics across products, and often limited explainability of AI predictions. Altogether, we aim to increase transparency of the current landscape of the clinician-AI interface and highlight the need to rigorously assess which strategies ultimately lead to the best clinical outcomes.
RESUMEN
Purpose To evaluate the use of ChatGPT as a tool to simplify answers to common questions about breast cancer prevention and screening. Materials and Methods In this retrospective, exploratory study, ChatGPT was requested to simplify responses to 25 questions about breast cancer to a sixth-grade reading level in March and August 2023. Simplified responses were evaluated for clinical appropriateness. All original and simplified responses were assessed for reading ease on the Flesch Reading Ease Index and for readability on five scales: Flesch-Kincaid Grade Level, Gunning Fog Index, Coleman-Liau Index, Automated Readability Index, and the Simple Measure of Gobbledygook (ie, SMOG) Index. Mean reading ease, readability, and word count were compared between original and simplified responses using paired t tests. McNemar test was used to compare the proportion of responses with adequate reading ease (score of 60 or greater) and readability (sixth-grade level). Results ChatGPT improved mean reading ease (original responses, 46 vs simplified responses, 70; P < .001) and readability (original, grade 13 vs simplified, grade 8.9; P < .001) and decreased word count (original, 193 vs simplified, 173; P < .001). Ninety-two percent (23 of 25) of simplified responses were considered clinically appropriate. All 25 (100%) simplified responses met criteria for adequate reading ease, compared with only two of 25 original responses (P < .001). Two of the 25 simplified responses (8%) met criteria for adequate readability. Conclusion ChatGPT simplified answers to common breast cancer screening and prevention questions by improving the readability by four grade levels, though the potential to produce incorrect information necessitates physician oversight when using this tool. Keywords: Mammography, Screening, Informatics, Breast, Education, Health Policy and Practice, Oncology, Technology Assessment Supplemental material is available for this article. © RSNA, 2023.
Asunto(s)
Neoplasias de la Mama , Alfabetización en Salud , Humanos , Femenino , Neoplasias de la Mama/diagnóstico , Neoplasias de la Mama/prevención & control , Detección Precoz del Cáncer , Estudios Retrospectivos , Atención Dirigida al PacienteRESUMEN
This Viewpoint discusses AI-generated clinical summaries and the necessity of transparent development of standards for their safe rollout.
Asunto(s)
Inteligencia Artificial , Registros Médicos , Alta del Paciente , Humanos , Exactitud de los DatosRESUMEN
Purpose To compare the effectiveness of weak supervision (ie, with examination-level labels only) and strong supervision (ie, with image-level labels) in training deep learning models for detection of intracranial hemorrhage (ICH) on head CT scans. Materials and Methods In this retrospective study, an attention-based convolutional neural network was trained with either local (ie, image level) or global (ie, examination level) binary labels on the Radiological Society of North America (RSNA) 2019 Brain CT Hemorrhage Challenge dataset of 21 736 examinations (8876 [40.8%] ICH) and 752 422 images (107 784 [14.3%] ICH). The CQ500 (436 examinations; 212 [48.6%] ICH) and CT-ICH (75 examinations; 36 [48.0%] ICH) datasets were employed for external testing. Performance in detecting ICH was compared between weak (examination-level labels) and strong (image-level labels) learners as a function of the number of labels available during training. Results On examination-level binary classification, strong and weak learners did not have different area under the receiver operating characteristic curve values on the internal validation split (0.96 vs 0.96; P = .64) and the CQ500 dataset (0.90 vs 0.92; P = .15). Weak learners outperformed strong ones on the CT-ICH dataset (0.95 vs 0.92; P = .03). Weak learners had better section-level ICH detection performance when more than 10 000 labels were available for training (average f1 = 0.73 vs 0.65; P < .001). Weakly supervised models trained on the entire RSNA dataset required 35 times fewer labels than equivalent strong learners. Conclusion Strongly supervised models did not achieve better performance than weakly supervised ones, which could reduce radiologist labor requirements for prospective dataset curation. Keywords: CT, Head/Neck, Brain/Brain Stem, Hemorrhage Supplemental material is available for this article. © RSNA, 2023 See also commentary by Wahid and Fuentes in this issue.
Asunto(s)
Aprendizaje Profundo , Humanos , Estudios Prospectivos , Estudios Retrospectivos , Hemorragias Intracraneales/diagnóstico por imagen , Tomografía Computarizada por Rayos XRESUMEN
A multidisciplinary physician team rated information provided by ChatGPT regarding breast pathologic diagnoses. ChatGPT responses were mostly appropriate regarding accuracy, consistency, definitions provided, and clinical significance conveyed. Responses were scored lower in terms of management recommendations provided, primarily related to low agreement with recommendations for high-risk lesions.
RESUMEN
GPT-4 outperformed a radiology domain-specific natural language processing model in classifying imaging findings from chest radiograph reports, both with and without predefined labels. Prompt engineering for context further improved performance. The findings indicate a role for large language models to accelerate artificial intelligence model development in radiology by automating data annotation.
Asunto(s)
Procesamiento de Lenguaje Natural , Radiografía Torácica , Humanos , Radiografía Torácica/métodos , Sistemas de Información RadiológicaRESUMEN
RATIONALE AND OBJECTIVES: With recent advancements in the power and accessibility of artificial intelligence (AI) Large Language Models (LLMs) patients might increasingly turn to these platforms to answer questions regarding radiologic examinations and procedures, despite valid concerns about the accuracy of information provided. This study aimed to assess the accuracy and completeness of information provided by the Bing Chatbot-a LLM powered by ChatGPT-on patient education for common radiologic exams. MATERIALS AND METHODS: We selected three common radiologic examinations and procedures: computed tomography (CT) abdomen, magnetic resonance imaging (MRI) spine, and bone biopsy. For each, ten questions were tested on the chatbot in two trials using three different chatbot settings. Two reviewers independently assessed the chatbot's responses for accuracy and completeness compared to an accepted online resource, radiologyinfo.org. RESULTS: Of the 360 reviews performed, 336 (93%) were rated "entirely correct" and 24 (7%) were "mostly correct," indicating a high level of reliability. Completeness ratings showed that 65% were "complete" and 35% were "mostly complete." The "More Creative" chatbot setting produced a higher proportion of responses rated "entirely correct" but there were otherwise no significant difference in ratings based on chatbot settings or exam types. The readability level was rated eighth-grade level. CONCLUSION: The Bing Chatbot provided accurate responses answering all or most aspects of the question asked of it, with responses tending to err on the side of caution for nuanced questions. Importantly, no responses were inaccurate or had potential to cause harm or confusion for the user. Thus, LLM chatbots demonstrate potential to enhance patient education in radiology and could be integrated into patient portals for various purposes, including exam preparation and results interpretation.
Asunto(s)
Inteligencia Artificial , Radiología , Humanos , Reproducibilidad de los Resultados , Educación del Paciente como Asunto , RadiografíaRESUMEN
Artificial intelligence (AI) is increasingly used in clinical practice for musculoskeletal imaging tasks, such as disease diagnosis and image reconstruction. AI applications in musculoskeletal imaging have focused primarily on radiography, CT, and MRI. Although musculoskeletal ultrasound stands to benefit from AI in similar ways, such applications have been relatively underdeveloped. In comparison with other modalities, ultrasound has unique advantages and disadvantages that must be considered in AI algorithm development and clinical translation. Challenges in developing AI for musculoskeletal ultrasound involve both clinical aspects of image acquisition and practical limitations in image processing and annotation. Solutions from other radiology subspecialties (e.g., crowdsourced annotations coordinated by professional societies), along with use cases (most commonly rotator cuff tendon tears and palpable soft-tissue masses), can be applied to musculoskeletal ultrasound to help develop AI. To facilitate creation of high-quality imaging datasets for AI model development, technologists and radiologists should focus on increasing uniformity in musculoskeletal ultrasound performance and increasing annotations of images for specific anatomic regions. This Expert Panel Narrative Review summarizes available evidence regarding AI's potential utility in musculoskeletal ultrasound and challenges facing its development. Recommendations for future AI advancement and clinical translation in musculoskeletal ultrasound are discussed.
Asunto(s)
Inteligencia Artificial , Tendones , Humanos , Ultrasonografía , Algoritmos , CabezaRESUMEN
OBJECTIVE: The purpose of this systematic review was to summarize the results of original research studies evaluating the characteristics and performance of deep learning models for detection of knee ligament and meniscus tears on MRI. MATERIALS AND METHODS: We searched PubMed for studies published as of February 2, 2022 for original studies evaluating development and evaluation of deep learning models for MRI diagnosis of knee ligament or meniscus tears. We summarized study details according to multiple criteria including baseline article details, model creation, deep learning details, and model evaluation. RESULTS: 19 studies were included with radiology departments leading the publications in deep learning development and implementation for detecting knee injuries via MRI. Among the studies, there was a lack of standard reporting and inconsistently described development details. However, all included studies reported consistently high model performance that significantly supplemented human reader performance. CONCLUSION: From our review, we found radiology departments have been leading deep learning development for injury detection on knee MRIs. Although studies inconsistently described DL model development details, all reported high model performance, indicating great promise for DL in knee MRI analysis.