No full-text available
To read the full-text of this research,
you can request a copy directly from the authors.
Large Language Models and Large Multimodal Models in Medical Imaging: A Primer for Physicians
Abstract
Large language models (LLMs) are poised to have a disruptive impact on health care. Numerous studies have demonstrated promising applications of LLMs in medical imaging, and this number will grow as LLMs further evolve into large multimodal models (LMMs) capable of processing both text and images. Given the substantial roles that LLMs and LMMs will have in health care, it is important for physicians to understand the underlying principles of these technologies so they can use them more effectively and responsibly and help guide their development. This article explains the key concepts behind the development and application of LLMs, including token embeddings, transformer networks, self-supervised pretraining, fine-tuning, and others. It also describes the technical process of creating LMMs and discusses use cases for both LLMs and LMMs in medical imaging.
ResearchGate has not been able to resolve any citations for this publication.
Traditional biomedical artificial intelligence (AI) models, designed for specific tasks or modalities, often exhibit limited flexibility in real-world deployment and struggle to utilize holistic information. Generalist AI holds the potential to address these limitations due to its versatility in interpreting different data types and generating tailored outputs for diverse needs. However, existing biomedical generalist AI solutions are typically heavyweight and closed source to researchers, practitioners and patients. Here, we describe BiomedGPT, the first open-source and lightweight vision–language foundation model, designed as a generalist capable of performing various biomedical tasks. BiomedGPT achieved state-of-the-art results in 16 out of 25 experiments while maintaining a computing-friendly model scale. We also conducted human evaluations to assess the capabilities of BiomedGPT in radiology visual question answering, report generation and summarization. BiomedGPT exhibits robust prediction ability with a low error rate of 3.8% in question answering, satisfactory performance with an error rate of 8.3% in writing complex radiology reports, and competitive summarization ability with a nearly equivalent preference score to human experts. Our method demonstrates that effective training with diverse data can lead to more practical biomedical AI for improving diagnosis and workflow efficiency.
The development of robust artificial intelligence models for echocardiography has been limited by the availability of annotated clinical data. Here, to address this challenge and improve the performance of cardiac imaging models, we developed EchoCLIP, a vision–language foundation model for echocardiography, that learns the relationship between cardiac ultrasound images and the interpretations of expert cardiologists across a wide range of patients and indications for imaging. After training on 1,032,975 cardiac ultrasound videos and corresponding expert text, EchoCLIP performs well on a diverse range of benchmarks for cardiac image interpretation, despite not having been explicitly trained for individual interpretation tasks. EchoCLIP can assess cardiac function (mean absolute error of 7.1% when predicting left ventricular ejection fraction in an external validation dataset) and identify implanted intracardiac devices (area under the curve (AUC) of 0.84, 0.92 and 0.97 for pacemakers, percutaneous mitral valve repair and artificial aortic valves, respectively). We also developed a long-context variant (EchoCLIP-R) using a custom tokenizer based on common echocardiography concepts. EchoCLIP-R accurately identified unique patients across multiple videos (AUC of 0.86), identified clinical transitions such as heart transplants (AUC of 0.79) and cardiac surgery (AUC 0.77) and enabled robust image-to-text search (mean cross-modal retrieval rank in the top 1% of candidate text reports). These capabilities represent a substantial step toward understanding and applying foundation models in cardiovascular imaging for preliminary interpretation of echocardiographic findings.
The accelerated adoption of digital pathology and advances in deep learning have enabled the development of robust models for various pathology tasks across a diverse array of diseases and patient cohorts. However, model training is often difficult due to label scarcity in the medical domain, and a model’s usage is limited by the specific task and disease for which it is trained. Additionally, most models in histopathology leverage only image data, a stark contrast to how humans teach each other and reason about histopathologic entities. We introduce CONtrastive learning from Captions for Histopathology (CONCH), a visual-language foundation model developed using diverse sources of histopathology images, biomedical text and, notably, over 1.17 million image–caption pairs through task-agnostic pretraining. Evaluated on a suite of 14 diverse benchmarks, CONCH can be transferred to a wide range of downstream tasks involving histopathology images and/or text, achieving state-of-the-art performance on histology image classification, segmentation, captioning, and text-to-image and image-to-text retrieval. CONCH represents a substantial leap over concurrent visual-language pretrained systems for histopathology, with the potential to directly facilitate a wide array of machine learning-based workflows requiring minimal or no further supervised fine-tuning.
Analyzing vast textual data and summarizing key information from electronic health records imposes a substantial burden on how clinicians allocate their time. Although large language models (LLMs) have shown promise in natural language processing (NLP) tasks, their effectiveness on a diverse range of clinical summarization tasks remains unproven. Here we applied adaptation methods to eight LLMs, spanning four distinct clinical summarization tasks: radiology reports, patient questions, progress notes and doctor–patient dialogue. Quantitative assessments with syntactic, semantic and conceptual NLP metrics reveal trade-offs between models and adaptation methods. A clinical reader study with 10 physicians evaluated summary completeness, correctness and conciseness; in most cases, summaries from our best-adapted LLMs were deemed either equivalent (45%) or superior (36%) compared with summaries from medical experts. The ensuing safety analysis highlights challenges faced by both LLMs and medical experts, as we connect errors to potential medical harm and categorize types of fabricated information. Our research provides evidence of LLMs outperforming medical experts in clinical text summarization across multiple tasks. This suggests that integrating LLMs into clinical workflows could alleviate documentation burden, allowing clinicians to focus more on patient care.
This study evaluated the ability of generative large language models (LLMs) to detect speech recognition errors in radiology reports. A dataset of 3233 CT and MRI reports was assessed by radiologists for speech recognition errors. Errors were categorized as clinically significant or not clinically significant. Performances of five generative LLMs-GPT-3.5-turbo, GPT-4, text-davinci-003, Llama-v2-70B-chat, and Bard-were compared in detecting these errors, using manual error detection as the reference standard. Prompt engineering was used to optimize model performance. GPT-4 demonstrated high accuracy in detecting clinically significant errors (precision, 76.9%; recall, 100%; F1 score, 86.9%) and not clinically significant errors (precision, 93.9%; recall, 94.7%; F1 score, 94.3%). Text-davinci-003 achieved F1 scores of 72% and 46.6% for clinically significant and not clinically significant errors, respectively. GPT-3.5-turbo obtained 59.1% and 32.2% F1 scores, while Llama-v2-70B-chat scored 72.8% and 47.7%. Bard showed the lowest accuracy, with F1 scores of 47.5% and 20.9%. GPT-4 effectively identified challenging errors of nonsense phrases and internally inconsistent statements. Longer reports, resident dictation, and overnight shifts were associated with higher error rates. In conclusion, advanced generative LLMs show potential for automatic detection of speech recognition errors in radiology reports. Keywords: CT, Large Language Model, Machine Learning, MRI, Natural Language Processing, Radiology Reports, Speech, Unsupervised Learning Supplemental material is available for this article.
Medical image segmentation is a critical component in clinical practice, facilitating accurate diagnosis, treatment planning, and disease monitoring. However, existing methods, often tailored to specific modalities or disease types, lack generalizability across the diverse spectrum of medical image segmentation tasks. Here we present MedSAM, a foundation model designed for bridging this gap by enabling universal medical image segmentation. The model is developed on a large-scale medical image dataset with 1,570,263 image-mask pairs, covering 10 imaging modalities and over 30 cancer types. We conduct a comprehensive evaluation on 86 internal validation tasks and 60 external validation tasks, demonstrating better accuracy and robustness than modality-wise specialist models. By delivering accurate and efficient segmentation across a wide spectrum of tasks, MedSAM holds significant potential to expedite the evolution of diagnostic tools and the personalization of treatment plans.
Large language models (LLMs) have demonstrated impressive capabilities, but the bar for clinical applications is high. Attempts to assess the clinical knowledge of models typically rely on automated evaluations based on limited benchmarks. Here, to address these limitations, we present MultiMedQA, a benchmark combining six existing medical question answering datasets spanning professional medicine, research and consumer queries and a new dataset of medical questions searched online, HealthSearchQA. We propose a human evaluation framework for model answers along multiple axes including factuality, comprehension, reasoning, possible harm and bias. In addition, we evaluate Pathways Language Model¹ (PaLM, a 540-billion parameter LLM) and its instruction-tuned variant, Flan-PaLM² on MultiMedQA. Using a combination of prompting strategies, Flan-PaLM achieves state-of-the-art accuracy on every MultiMedQA multiple-choice dataset (MedQA³, MedMCQA⁴, PubMedQA⁵ and Measuring Massive Multitask Language Understanding (MMLU) clinical topics⁶), including 67.6% accuracy on MedQA (US Medical Licensing Exam-style questions), surpassing the prior state of the art by more than 17%. However, human evaluation reveals key gaps. To resolve this, we introduce instruction prompt tuning, a parameter-efficient approach for aligning LLMs to new domains using a few exemplars. The resulting model, Med-PaLM, performs encouragingly, but remains inferior to clinicians. We show that comprehension, knowledge recall and reasoning improve with model scale and instruction prompt tuning, suggesting the potential utility of LLMs in medicine. Our human evaluations reveal limitations of today’s models, reinforcing the importance of both evaluation frameworks and method development in creating safe, helpful LLMs for clinical applications.
Artificial intelligence (AI) models for decision support have been developed for clinical settings such as radiology, but little work evaluates the potential impact of such systems. In this study, physicians received chest X-rays and diagnostic advice, some of which
was inaccurate, and were asked to evaluate advice quality and make diagnoses. All advice was generated by human experts, but some was labeled as coming from an AI system. As a group, radiologists rated advice as lower quality when it appeared to come from an AI system; physicians with less task-expertise did not. Diagnostic accuracy was significantly worse when participants received inaccurate advice, regardless of the purported source. This work raises important considerations for how advice, AI and non-AI, should be deployed in clinical environments.
We propose two novel model architectures for computing continuous vector
representations of words from very large data sets. The quality of these
representations is measured in a word similarity task, and the results are
compared to the previously best performing techniques based on different types
of neural networks. We observe large improvements in accuracy at much lower
computational cost, i.e. it takes less than a day to learn high quality word
vectors from a 1.6 billion words data set. Furthermore, we show that these
vectors provide state-of-the-art performance on our test set for measuring
syntactic and semantic word similarities.
A new recurrent neural network based language model (RNN LM) with applications to speech recognition is presented. Results indicate that it is possible to obtain around 50% reduction of perplexity by using mixture of several RNN LMs, compared to a state of the art backoff language model. Speech recognition experiments show around 18% reduction of word error rate on the Wall Street Journal task when comparing models trained on the same amount of data, and around 5% on the much harder NIST RT05 task, even when the backoff model is trained on much more data than the RNN LM. We provide ample empirical evidence to suggest that connectionist language models are superior to standard n-gram techniques, except their high computational (training) complexity.
Learning to store information over extended time intervals by recurrent backpropagation takes a very long time, mostly because of insufficient, decaying error backflow. We briefly review Hochreiter's (1991) analysis of this problem, then address it by introducing a novel, efficient, gradient based method called long short-term memory (LSTM). Truncating the gradient where this does not do harm, LSTM can learn to bridge minimal time lags in excess of 1000 discrete-time steps by enforcing constant error flow through constant error carousels within special units. Multiplicative gate units learn to open and close access to the constant error flow. LSTM is local in space and time; its computational complexity per time step and weight is O. 1. Our experiments with artificial data involve local, distributed, real-valued, and noisy pattern representations. In comparisons with real-time recurrent learning, back propagation through time, recurrent cascade correlation, Elman nets, and neural sequence chunking, LSTM leads to many more successful runs, and learns much faster. LSTM also solves complex, artificial long-time-lag tasks that have never been solved by previous recurrent network algorithms.
Human evaluations of machine translation are extensive but expensive. Human evaluations can take months to finish and involve human labor that can not be reused.
A goal of statistical language modeling is to learn the joint probability function of sequences of words. This is intrinsically difficult because of the curse of dimensionality: we propose to fight it with its own weapons.
Introduction: The potential of Large Language Models (LLMs) in enhancing a variety of natural language tasks in clinical fields includes medical imaging reporting. This pilot study examines the efficacy of a retrieval-augmented LLM system considering zero-shot learning capability of LLMs, integrated with a comprehensive PET reading reports database, in improving referring previous reports and decision-making.
Methods: We developed a custom LLM framework enhanced with retrieval capabilities, leveraging a database encompassing nine years of PET imaging reports from a single center. The system employs vector space embedding of the reports database to facilitate retrieval based on similarity metrics. Queries prompt the system to retrieve embedded vectors, generating context-based answers and identifying similar cases or differential diagnoses from the historical reports database.
Results: The system efficiently organized embedded vectors from PET reading reports, showing that imaging reports were accurately clustered within the embedded vector space according to the diagnosis or PET study type. Based on this system, a proof-of-concept chatbot was developed and showed the framework's potential in referencing reports of previous similar cases and identifying exemplary cases for various purposes. Additionally, it demonstrated the capability to offer differential diagnoses, leveraging the vast database to enhance the completeness and precision of generated reports.
Conclusions: The integration of a retrieval-augmented LLM with a large database of PET imaging reports represents an advancement in medical reporting within nuclear medicine. By providing tailored, data-driven insights, the system not only improves the relevance of PET report generation but also supports enhanced decision-making and educational opportunities. This study underscores the potential of advanced AI tools in transforming medical imaging reporting practices.
Purpose
To assess the performance of a local open-source large language model (LLM) in various information extraction tasks from real-life emergency brain MRI reports.
Materials and Methods
All consecutive emergency brain MRI reports written in 2022 from a French quaternary center were retrospectively reviewed. Two radiologists identified MRI scans that were performed in the emergency department for headaches. Four radiologists scored the reports’ conclusions as either normal or abnormal. Abnormalities were labeled as either headache-causing or incidental. Vicuna (LMSYS Org), an open-source LLM, performed the same tasks. Vicuna’s performance metrics were evaluated using the radiologists’ consensus as the reference standard.
Results
Among the 2398 reports during the study period, radiologists identified 595 that included headaches in the indication (median age of patients, 35 years [IQR, 26–51 years]; 68% [403 of 595] women). A positive finding was reported in 227 of 595 (38%) cases, 136 of which could explain the headache. The LLM had a sensitivity of 98.0% (95% CI: 96.5, 99.0) and specificity of 99.3% (95% CI: 98.8, 99.7) for detecting the presence of headache in the clinical context, a sensitivity of 99.4% (95% CI: 98.3, 99.9) and specificity of 98.6% (95% CI: 92.2, 100.0) for the use of contrast medium injection, a sensitivity of 96.0% (95% CI: 92.5, 98.2) and specificity of 98.9% (95% CI: 97.2, 99.7) for study categorization as either normal or abnormal, and a sensitivity of 88.2% (95% CI: 81.6, 93.1) and specificity of 73% (95% CI: 62, 81) for causal inference between MRI findings and headache.
Conclusion
An open-source LLM was able to extract information from free-text radiology reports with excellent accuracy without requiring further training.
Keywords: Large Language Model (LLM), Generative Pretrained Transformers (GPT), Open Source, Information Extraction, Report, Brain, MRI
Supplemental material is available for this article.
Published under a CC BY 4.0 license.
See also the commentary by Akinci D’Antonoli and Bluethgen in this issue.
Background Errors in radiology reports may occur because of resident-to-attending discrepancies, speech recognition inaccuracies, and large workload. Large language models, such as GPT-4 (ChatGPT; OpenAI), may assist in generating reports. Purpose To assess effectiveness of GPT-4 in identifying common errors in radiology reports, focusing on performance, time, and cost-efficiency. Materials and Methods In this retrospective study, 200 radiology reports (radiography and cross-sectional imaging [CT and MRI]) were compiled between June 2023 and December 2023 at one institution. There were 150 errors from five common error categories (omission, insertion, spelling, side confusion, and other) intentionally inserted into 100 of the reports and used as the reference standard. Six radiologists (two senior radiologists, two attending physicians, and two residents) and GPT-4 were tasked with detecting these errors. Overall error detection performance, error detection in the five error categories, and reading time were assessed using Wald χ2 tests and paired-sample t tests. Results GPT-4 (detection rate, 82.7%;124 of 150; 95% CI: 75.8, 87.9) matched the average detection performance of radiologists independent of their experience (senior radiologists, 89.3% [134 of 150; 95% CI: 83.4, 93.3]; attending physicians, 80.0% [120 of 150; 95% CI: 72.9, 85.6]; residents, 80.0% [120 of 150; 95% CI: 72.9, 85.6]; P value range, .522-.99). One senior radiologist outperformed GPT-4 (detection rate, 94.7%; 142 of 150; 95% CI: 89.8, 97.3; P = .006). GPT-4 required less processing time per radiology report than the fastest human reader in the study (mean reading time, 3.5 seconds ± 0.5 [SD] vs 25.1 seconds ± 20.1, respectively; P < .001; Cohen d = -1.08). The use of GPT-4 resulted in lower mean correction cost per report than the most cost-efficient radiologist (0.42 ± 0.41; P < .001; Cohen d = -1.12). Conclusion The radiology report error detection rate of GPT-4 was comparable with that of radiologists, potentially reducing work hours and cost. © RSNA, 2024 See also the editorial by Forman in this issue.
Radiology narrative reports often describe characteristics of a patient’s disease, including its location, size, and shape. Motivated by the recent success of multimodal learning, we hypothesized that this descriptive text could guide medical image analysis algorithms. We proposed a novel vision-language model, ConTEXTual Net, for the task of pneumothorax segmentation on chest radiographs. ConTEXTual Net extracts language features from physician-generated free-form radiology reports using a pre-trained language model. We then introduced cross-attention between the language features and the intermediate embeddings of an encoder-decoder convolutional neural network to enable language guidance for image analysis. ConTEXTual Net was trained on the CANDID-PTX dataset consisting of 3196 positive cases of pneumothorax with segmentation annotations from 6 different physicians as well as clinical radiology reports. Using cross-validation, ConTEXTual Net achieved a Dice score of 0.716±0.016, which was similar to the degree of inter-reader variability (0.712±0.044) computed on a subset of the data. It outperformed vision-only models (Swin UNETR: 0.670±0.015, ResNet50 U-Net: 0.677±0.015, GLoRIA: 0.686±0.014, and nnUNet 0.694±0.016) and a competing vision-language model (LAVT: 0.706±0.009). Ablation studies confirmed that it was the text information that led to the performance gains. Additionally, we show that certain augmentation methods degraded ConTEXTual Net’s segmentation performance by breaking the image-text concordance. We also evaluated the effects of using different language models and activation functions in the cross-attention module, highlighting the efficacy of our chosen architectural design.
The “Impression” section of a radiology report is a critical basis for communication between radiologists and other physicians. Typically written by radiologists, this part is derived from the “Findings” section, which can be laborious and error-prone. Although deep-learning based models, such as BERT, have achieved promising results in Automatic Impression Generation (AIG), such models often require substantial amounts of medical data and have poor generalization performance. Recently, Large Language Models (LLMs) like ChatGPT have shown strong generalization capabilities and performance, but their performance in specific domains, such as radiology, remains under-investigated and potentially limited. To address this limitation, we propose ImpressionGPT, leveraging the contextual learning capabilities of LLMs through our dynamic prompt and iterative optimization algorithm to accomplish the AIG task. ImpressionGPT initially employs a small amount of domain-specific data to create a dynamic prompt, extracting contextual semantic information closely related to the test data. Subsequently, the iterative optimization algorithm automatically evaluates the output of LLMs and provides optimization suggestions, continuously refining the output results. The proposed ImpressionGPT model achieves superior performance of AIG task on both MIMIC-CXR and OpenI datasets without requiring additional training data or fine-tuning the LLMs. This work presents a paradigm for localizing LLMs that can be applied in a wide range of similar application scenarios, bridging the gap between general-purpose LLMs and the specific language processing needs of various domains.
Purpose
To determine if fine-tuned large language models (LLMs) can generate accurate, personalized impressions for whole-body PET reports.
Materials and Methods
Twelve language models were trained on a corpus of PET reports using the teacher-forcing algorithm, with the report findings as input and the clinical impressions as reference. An extra input token encodes the reading physician’s identity, allowing models to learn physician-specific reporting styles. Our corpus comprised 37,370 retrospective PET reports collected from our institution between 2010 and 2022. To identify the best LLM, 30 evaluation metrics were benchmarked against quality scores from two nuclear medicine (NM) physicians, with the most aligned metrics selecting the model for expert evaluation. In a subset of data, model-generated impressions and original clinical impressions were assessed by three NM physicians according to 6 quality dimensions (3-point scale) and an overall utility score (5-point scale). Each physician reviewed 12 of their own reports and 12 reports from other physicians. Bootstrap resampling was used for statistical analysis.
Results
Of all evaluation metrics, domain-adapted BARTScore and PEGASUSScore showed the highest Spearman’s ρ correlations (ρ=0.568 and 0.563) with physician preferences. Based on these metrics, the fine-tuned PEGASUS model was selected as the top LLM. When physicians reviewed PEGASUS-generated impressions in their own style, 89% were considered clinically acceptable, with a mean utility score of 4.08 out of 5. Physicians rated these personalized impressions as comparable in overall utility to the impressions dictated by other physicians (4.03, P=0.41).
Conclusion
Personalized impressions generated by PEGASUS were clinically useful, highlighting its potential to expedite PET reporting.
Purpose:
To evaluate the impact of domain adaptation on the performance of language models in predicting five-point Deauville scores on the basis of clinical fluorine 18 fluorodeoxyglucose PET/CT reports.
Materials and methods:
The authors retrospectively retrieved 4542 text reports and images for fluorodeoxyglucose PET/CT lymphoma examinations from 2008 to 2018 in the University of Wisconsin-Madison institutional clinical imaging database. Of these total reports, 1664 had Deauville scores that were extracted from the reports and served as training labels. The bidirectional encoder representations from transformers (BERT) model and initialized BERT models BioClinicalBERT, RadBERT, and RoBERTa were adapted to the nuclear medicine domain by pretraining using masked language modeling. These domain-adapted models were then compared with the non-domain-adapted versions on the task of five-point Deauville score prediction. The language models were compared against vision models, multimodal vision-language models, and a nuclear medicine physician, with sevenfold Monte Carlo cross-validation. Means and SDs for accuracy are reported, with P values from paired t testing.
Results:
Domain adaptation improved the performance of all language models (P = .01). For example, BERT improved from 61.3% ± 2.9 (SD) five-class accuracy to 65.7% ± 2.2 (P = .01) following domain adaptation. Domain-adapted RoBERTa (named DA RoBERTa) performed best, achieving 77.4% ± 3.4 five-class accuracy; this model performed similarly to its multimodal counterpart (named Multimodal DA RoBERTa) (77.2% ± 3.2) and outperformed the best vision-only model (48.1% ± 3.5, P ≤ .001). A physician given the task on a subset of the data had a five-class accuracy of 66%.
Conclusion:
Domain adaptation improved the performance of large language models in predicting Deauville scores in PET/CT reports.Keywords Lymphoma, PET, PET/CT, Transfer Learning, Unsupervised Learning, Convolutional Neural Network (CNN), Nuclear Medicine, Deauville, Natural Language Processing, Multimodal Learning, Artificial Intelligence, Machine Learning, Language Modeling Supplemental material is available for this article. © RSNA, 2023See also the commentary by Abajian in this issue.
Background The latest large language models (LLMs) solve unseen problems via user-defined text prompts without the need for retraining, offering potentially more efficient information extraction from free-text medical records than manual annotation. Purpose To compare the performance of the LLMs ChatGPT and GPT-4 in data mining and labeling oncologic phenotypes from free-text CT reports on lung cancer by using user-defined prompts. Materials and Methods This retrospective study included patients who underwent lung cancer follow-up CT between September 2021 and March 2023. A subset of 25 reports was reserved for prompt engineering to instruct the LLMs in extracting lesion diameters, labeling metastatic disease, and assessing oncologic progression. This output was fed into a rule-based natural language processing pipeline to match ground truth annotations from four radiologists and derive performance metrics. The oncologic reasoning of LLMs was rated on a five-point Likert scale for factual correctness and accuracy. The occurrence of confabulations was recorded. Statistical analyses included Wilcoxon signed rank and McNemar tests. Results On 424 CT reports from 424 patients (mean age, 65 years ± 11 [SD]; 265 male), GPT-4 outperformed ChatGPT in extracting lesion parameters (98.6% vs 84.0%, P < .001), resulting in 96% correctly mined reports (vs 67% for ChatGPT, P < .001). GPT-4 achieved higher accuracy in identification of metastatic disease (98.1% [95% CI: 97.7, 98.5] vs 90.3% [95% CI: 89.4, 91.0]) and higher performance in generating correct labels for oncologic progression (F1 score, 0.96 [95% CI: 0.94, 0.98] vs 0.91 [95% CI: 0.89, 0.94]) (both P < .001). In oncologic reasoning, GPT-4 had higher Likert scale scores for factual correctness (4.3 vs 3.9) and accuracy (4.4 vs 3.3), with a lower rate of confabulation (1.7% vs 13.7%) than ChatGPT (all P < .001). Conclusion When using user-defined prompts, GPT-4 outperformed ChatGPT in extracting oncologic phenotypes from free-text CT reports on lung cancer and demonstrated better oncologic reasoning with fewer confabulations. © RSNA, 2023 Supplemental material is available for this article. See also the editorial by Hafezi-Nejad and Trivedi in this issue.
Background The recent release of large language models (LLMs) for public use, such as ChatGPT and Google Bard, has opened up a multitude of potential benefits as well as challenges. Purpose To evaluate and compare the accuracy and consistency of responses generated by publicly available ChatGPT-3.5 and Google Bard to non-expert questions related to lung cancer prevention, screening, and terminology commonly used in radiology reports based on the recommendation of Lung Imaging Reporting and Data System (Lung-RADS) v2022 from American College of Radiology and Fleischner society. Materials and Methods Forty of the exact same questions were created and presented to ChatGPT-3.5 and Google Bard experimental version as well as Bing and Google search engines by three different authors of this paper. Each answer was reviewed by two radiologists for accuracy. Responses were scored as correct, partially correct, incorrect, or unanswered. Consistency was also evaluated among the answers. Here, consistency was defined as the agreement between the three answers provided by ChatGPT-3.5, Google Bard experimental version, Bing, and Google search engines regardless of whether the concept conveyed was correct or incorrect. The accuracy among different tools were evaluated using Stata. Results ChatGPT-3.5 answered 120 questions with 85 (70.8%) correct, 14 (11.7%) partially correct, and 21 (17.5%) incorrect. Google Bard did not answer 23 (19.1%) questions. Among the 97 questions answered by Google Bard, 62 (51.7%) were correct, 11 (9.2%) were partially correct, and 24 (20%) were incorrect. Bing answered 120 questions with 74 (61.7%) correct, 13 (10.8%) partially correct, and 33 (27.5%) incorrect. Google search engine answered 120 questions with 66 (55%) correct, 27 (22.5%) partially correct, and 27 (22.5%) incorrect. The ChatGPT-3.5 is more likely to provide correct or partially answer than Google Bard, approximately by 1.5 folds (OR = 1.55, P = 0.004). ChatGPT-3.5 and Google search engine were more likely to be consistent than Google Bard by approximately 7 and 29 folds (OR = 6.65, P = 0.002 for ChatGPT and OR = 28.83, P = 0.002 for Google search engine, respectively). Conclusion Although ChatGPT-3.5 had a higher accuracy in comparison with the other tools, neither ChatGPT nor Google Bard, Bing and Google search engines answered all questions correctly and with 100% consistency.
GPT-4 automates the transformation of various free-text radiology reports into structured templates with minor effort, overcoming the challenges of implementing structured reporting while improving standardization and data extraction for research.
Purpose:
To investigate if tailoring a transformer-based language model to radiology is beneficial for radiology natural language processing (NLP) applications.
Materials and methods:
This retrospective study presents a family of bidirectional encoder representations from transformers (BERT)-based language models adapted for radiology, named RadBERT. Transformers were pretrained with either 2.16 or 4.42 million radiology reports from U.S. Department of Veterans Affairs health care systems nationwide on top of four different initializations (BERT-base, Clinical-BERT, robustly optimized BERT pretraining approach [RoBERTa], and BioMed-RoBERTa) to create six variants of RadBERT. Each variant was fine-tuned for three representative NLP tasks in radiology: (a) abnormal sentence classification: models classified sentences in radiology reports as reporting abnormal or normal findings; (b) report coding: models assigned a diagnostic code to a given radiology report for five coding systems; and (c) report summarization: given the findings section of a radiology report, models selected key sentences that summarized the findings. Model performance was compared by bootstrap resampling with five intensively studied transformer language models as baselines: BERT-base, BioBERT, Clinical-BERT, BlueBERT, and BioMed-RoBERTa.
Results:
For abnormal sentence classification, all models performed well (accuracies above 97.5 and F1 scores above 95.0). RadBERT variants achieved significantly higher scores than corresponding baselines when given only 10% or less of 12 458 annotated training sentences. For report coding, all variants outperformed baselines significantly for all five coding systems. The variant RadBERT-BioMed-RoBERTa performed the best among all models for report summarization, achieving a Recall-Oriented Understudy for Gisting Evaluation-1 score of 16.18 compared with 15.27 by the corresponding baseline (BioMed-RoBERTa, P < .004).
Conclusion:
Transformer-based language models tailored to radiology had improved performance of radiology NLP tasks compared with baseline transformer language models.Keywords: Translation, Unsupervised Learning, Transfer Learning, Neural Networks, Informatics Supplemental material is available for this article. © RSNA, 2022See also commentary by Wiggins and Tejani in this issue.
An important need exists for strategies to perform rigorous objective clinical-task-based evaluation of artificial intelligence (AI) algorithms for nuclear medicine. To address this need, we propose a four-class framework to evaluate AI algorithms for promise, technical task-specific efficacy, clinical decision making, and post-deployment efficacy. We provide best practices to evaluate AI algorithms for each of these classes. Each class of evaluation yields a claim that provides a descriptive performance of the AI algorithm. Key best practices are tabulated as the RELAINCE (Recommendations for EvaLuation of AI for NuClear medicinE) guidelines. The report was prepared by the Society of Nuclear Medicine and Molecular Imaging AI taskforce Evaluation team, which consisted of nuclear-medicine physicians, physicists, computational imaging scientists, and representatives from industry and regulatory agencies.
Large pre-trained language models have been shown to store factual knowledge in their parameters, and achieve state-of-the-art results when fine-tuned on downstream NLP tasks. However, their ability to access and precisely manipulate knowledge is still limited, and hence on knowledge-intensive tasks, their performance lags behind task-specific architectures. Additionally, providing provenance for their decisions and updating their world knowledge remain open research problems. Pre-trained models with a differentiable access mechanism to explicit non-parametric memory can overcome this issue, but have so far been only investigated for extractive downstream tasks. We explore a general-purpose fine-tuning recipe for retrieval-augmented generation (RAG) -- models which combine pre-trained parametric and non-parametric memory for language generation. We introduce RAG models where the parametric memory is a pre-trained seq2seq model and the non-parametric memory is a dense vector index of Wikipedia, accessed with a pre-trained neural retriever. We compare two RAG formulations, one which conditions on the same retrieved passages across the whole generated sequence, the other can use different passages per token. We fine-tune and evaluate our models on a wide range of knowledge-intensive NLP tasks and set the state-of-the-art on three open domain QA tasks, outperforming parametric seq2seq models and task-specific retrieve-and-extract architectures. For language generation tasks, we find that RAG models generate more specific, diverse and factual language than a state-of-the-art parametric-only seq2seq baseline.
We propose two novel model architectures for computing continuous vector representations of words from very large data sets. The quality of these representations is measured in a word similarity task, and the results are compared to the previously best performing techniques based on different types of neural networks. We observe large improvements in accuracy at much lower computational cost, i.e. it takes less than a day to learn high quality word vectors from a 1.6 billion words data set. Furthermore, we show that these vectors provide state-of-the-art performance on our test set for measuring syntactic and semantic word similarities.
The dominant sequence transduction models are based on complex recurrent or convolutional neural networks in an encoder-decoder configuration. The best performing models also connect the encoder and decoder through an attention mechanism. We propose a new simple network architecture, the Transformer, based solely on attention mechanisms, dispensing with recurrence and convolutions entirely. Experiments on two machine translation tasks show these models to be superior in quality while being more parallelizable and requiring significantly less time to train. Our model achieves 28.4 BLEU on the WMT 2014 English-to-German translation task, improving over the existing best results, including ensembles by over 2 BLEU. On the WMT 2014 English-to-French translation task, our model establishes a new single-model state-of-the-art BLEU score of 41.0 after training for 3.5 days on eight GPUs, a small fraction of the training costs of the best models from the literature. We show that the Transformer generalizes well to other tasks by applying it successfully to English constituency parsing both with large and limited training data.
ROUGE stands for Recall-Oriented Understudy for Gisting Evaluation. It includes measures to automatically determine the quality of a summary by comparing it to other (ideal) summaries created by humans. The measures count the number of overlapping units such as n-gram, word sequences, and word pairs between the computer-generated summary to be evaluated and the ideal summaries created by humans. This paper introduces four different ROUGE measures: ROUGE-N, ROUGE-L, ROUGE-W, and ROUGE-S included in the ROUGE summarization evaluation package and their evaluations. Three of them have been used in the Document Understanding Conference (DUC) 2004, a large-scale sum- marization evaluation sponsored by NIST.
We survey the most widely-used algorithms for smoothing models for language n -gram modeling. We then present an extensive empirical comparison of several of these smoothing techniques, including those described by Jelinek and Mercer (1980); Katz (1987); Bell, Cleary and Witten (1990); Ney, Essen and Kneser (1994), and Kneser and Ney (1995). We investigate how factors such as training data size, training corpus (e.g. Brown vs. Wall Street Journal), count cutoffs, and n -gram order (bigram vs. trigram) affect the relative performance of these methods, which is measured through the cross-entropy of test data. We find that these factors can significantly affect the relative performance of models, with the most significant factor being training data size. Since no previous comparisons have examined these factors systematically, this is the first thorough characterization of the relative performance of various algorithms. In addition, we introduce methodologies for analyzing smoothing algorithm efficacy in detail, and using these techniques we motivate a novel variation of Kneser–Ney smoothing that consistently outperforms all other algorithms evaluated. Finally, results showing that improved language model smoothing leads to improved speech recognition performance are presented.
Eliza is a program operating within the MAC time-sharing system at MIT which makes certain kinds of natural language conversation between man and computer possible. Input sentences are analyzed on the basis of decomposition rules which are triggered by key words appearing in the input text. Responses are generated by reassembly rules associated with selected decomposition rules. The fundamental technical problems with which Eliza is concerned are: (1) the identification of key words, (2) the discovery of minimal context, (3) the choice of appropriate transformations, (4) generation of responses in the absence of key words, and (5) the provision of an editing capability for Eliza scripts. A discussion of some psychological issues relevant to the Eliza approach as well as of future developments concludes the paper. 9 references.
A generalist learner for multifaceted medical image interpretation
- H-Y Adithan S Zhou
- J N Acosta
- Topol
- P Rajpurkar
Zhou H-Y, Adithan S, Acosta JN, Topol EJ, Rajpurkar P. A generalist learner for
multifaceted medical image interpretation. arXiv website. https://arxiv.org/abs/
2405.07988. Published May 13, 2024. Accessed December 31, 2024.
Direct preference optimization: your language model is secretly a reward model
- R Rafailov
- Sharma
- S Mitchell E Ermon
- C D Manning
- C Finn
Rafailov R, Sharma A, Mitchell E, Ermon S, Manning CD, Finn C. Direct preference optimization: your language model is secretly a reward model. arXiv website.
https://arxiv.org/abs/2305.18290. Published May 29, 2023. Revised July 29, 2024.
Accessed August 5, 2024.
Advancing multimodal medical capabilities of Gemini
- A Yang L Xu S Sellergren
Improving language understanding by generative pre-training. OpenAI website
- T Radford A Narasimhan K Salimans
- I Sutskever
Radford A, Narasimhan K, Salimans T, Sutskever I. Improving language understanding by generative pre-training. OpenAI website. https://openai.com/research/
language-unsupervised. Published June 11, 2018. Accessed December 31, 2024.
Language models are few-shot learners
- T B Brown
- N Mann B Ryder
Brown TB, Mann B, Ryder N, et al. Language models are few-shot learners. In:
Proceedings of the 34th International Conference on Neural Information Processing Systems. NIPS '20. Curran Associates Inc.; 2020:1877-1901.
Zero-shot text-to-image generation
- Ramesh A Pavlov M Goh