Jun Kanzawa

Purpose This study examined whether task-based functional MRI (fMRI) can provide metrics of local brain activity and hemodynamics typically derived from resting-state fMRI (rsfMRI). Methods Two publicly open datasets from healthy individuals and brain tumor patients were retrospectively used to compare amplitude of low-frequency fluctuations (ALFF) and global signal metrics, namely, global signal correlation (GSC), global signal cross-correlation (GSXC), and global signal time delay (GSTD) derived from motor and language fMRI with those from a reference rsfMRI scan (rest-1). Factors influencing the correlations between each measurement were examined using a mixed-effects model with post-hoc pairwise comparisons. Results Our study included 50 healthy subjects (mean age: 29 ± 3 years; 32 women) and 38 patients (45 ± 14 years; 24 men). Significant correlations (p < 0.001) were found for all metrics between those derived from motor, language, or a second rest scan (rest-2) against rest-1 scans for both healthy subjects and patients. Both fMRI type and motion magnitude affected the measurements of healthy subjects’ GSC (fMRI type: F [2, 69.1] = 15.4, p < 0.001; motion: F [1, 123.1] = 8.5, p = 0.004), GSXC (fMRI type: F [2, 69.5] = 18.9, p < 0.001; motion: F [1, 123.2] = 9.89, p = 0.002), and GSTD (fMRI type: F [2, 58.7] = 9.89, p < 0.001; motion: F [1, 123.2] = 15.26, p < 0.001). FMRI type by motion interaction was significant for ALFF (F [2, 65.9] = 6.13, p = 0.004). The effects were less pronounced and observed only for GSC (F [2, 40.4] = 6.1, p = 0.005) and GSXC (F [2, 41.2] = 6.4, p = 0.004) in the clinical dataset. Conclusion Repurposing existing task-based functional MRI data for evaluating local brain activities and hemodynamics is feasible in the entire brain in healthy and brain tumor subjects.

Background and Purpose: This study aims to characterize hemodynamic alterations in gliomas by analyzing global signal metrics and amplitude of low‐frequency fluctuations (ALFF). Methods: We analyzed resting‐state functional MRI data from an open dataset. A total of 29 glioma patients (17 men; mean age: 44.2 ± 13.9 years; 16 high‐grade) were included. Global signal cross‐correlation (GSXC), representing the maximum cross‐correlation value between each voxel's time series and the average gray matter signal of the contralateral hemisphere, and ALFF were calculated. Regions of interest for tumor were defined and mirrored to the contralateral side for comparison. Linear mixed‐effects models were used to examine effects of tumor grade and measurement side on GSXC and ALFF. Post hoc analyses compared tumor versus contralateral regions. Results: GSXC was significantly reduced in tumor regions compared to the contralateral hemisphere, regardless of tumor grade (low‐grade: 0.31 ± 0.21 vs. 0.45 ± 0.16, p = 0.005; high‐grade: 0.39 ± 0.17 vs. 0.55 ± 0.13, p < 0.001). ALFF was significantly lower in low‐grade tumors compared to the contralateral side (0.65 ± 0.15 vs. 0.77 ± 0.14, p = 0.005), while no significant difference was found in high‐grade tumors (0.80 ± 0.18 vs. 0.77 ± 0.10, p = 0.46). Conclusions: Our findings suggest that GSXC and ALFF provide complementary pathological information on glioma. GSXC indicates impaired cerebrovascular reactivity due to gliovascular uncoupling, whereas ALFF shows an increase in high‐grade gliomas likely reflecting greater total vascularity.

Background: Large Language Models (LLMs) show promise in medical diagnosis, but their performance varies with prompting. Recent studies suggest that modifying prompts may enhance diagnostic capabilities. Objective: This study aimed to test whether a prompting approach that aligns with general clinical reasoning methodology—specifically, separating processes of summarizing clinical information and making diagnoses based on the summary instead of one-step processing—can enhance LLM’s medical diagnostic capabilities Methods: 322 quiz questions from Radiology’s Diagnosis Please cases (1998-2023) were used. We employed Claude 3.5 Sonnet, a state-of-the-art LLM, to compare three approaches: 1) Conventional zero-shot chain-of-thought prompt, as a baseline, 2) two-step approach: LLM organizes patient history and imaging findings, then provides diagnoses, and 3) Summary-only approach: Using only the LLM-generated summary for diagnoses. Results: The two-step approach significantly outperformed both baseline and summary-only methods in diagnosis accuracy, as determined by McNemar tests. Primary diagnosis accuracy was 60.6% for the two-step approach, compared to 56.5% for baseline (p=0.042) and 56.3% for summary-only (p=0.035). For the top three diagnoses, accuracy was 70.5%, 66.5%, and 65.5% respectively (p=0.005 for baseline, p=0.008 for summary-only). No significant differences were observed between baseline and summary-only approaches. Conclusion: Our results indicate that a structured clinical reasoning approach enhances LLM’s diagnostic accuracy. This method shows potential as a valuable tool for deriving diagnoses from free-text clinical information. The approach aligns well with established clinical reasoning processes, suggesting its potential applicability in real-world clinical settings.

Background Large language models (LLMs) show promise in radiological diagnosis, but their performance may be affected by the context of the cases presented. Purpose To investigate how providing information about prior probabilities influences the diagnostic performance of an LLM in radiological quiz cases. Materials and Methods We analyzed 322 consecutive cases from Radiology’s “Diagnosis Please” quiz using Claude 3.5 Sonnet under three conditions: without context (Condition 1), informed as quiz cases (Condition 2), and presented as primary care cases (Condition 3). Diagnostic accuracy was compared using McNemar’s test. Results The overall accuracy rate significantly improved in Condition 2 compared to Condition 1 (70.2% vs. 64.9%, p=0.029). Conversely, the accuracy rate significantly decreased in Condition 3 compared to Condition 1 (59.9% vs. 70.2%, pConclusion Providing context about prior probabilities significantly affects the diagnostic performance of the LLM in radiological cases. This suggests that LLMs may incorporate Bayesian-like principles in their diagnostic approach, highlighting the potential for optimizing LLM’s performance in clinical settings by providing relevant contextual information. Key Results LLM’s overall accuracy improved from 64.9% to 70.2% when informed about quiz case nature (p=0.029). LLM’s overall accuracy decreased to 59.9% when presented with incorrect primary care context (pSummary Statement Providing context about prior probabilities significantly influences LLM’s diagnostic performance in radiological cases, suggesting potential for optimizing LLM use in clinical practice through contextual information.

Early detection of patients with impending bone metastasis is crucial for prognosis improvement. This study aimed to investigate the feasibility of a fine-tuned, locally run large language model (LLM) in extracting patients with bone metastasis in unstructured Japanese radiology report and to compare its performance with manual annotation. This retrospective study included patients with “metastasis” in radiological reports (April 2018–January 2019, August–May 2022, and April–December 2023 for training, validation, and test datasets of 9559, 1498, and 7399 patients, respectively). Radiologists reviewed the clinical indication and diagnosis sections of the radiological report (used as input data) and classified them into groups 0 (no bone metastasis), 1 (progressive bone metastasis), and 2 (stable or decreased bone metastasis). The data for group 0 was under-sampled in training and test datasets due to group imbalance. The best-performing model from the validation set was subsequently tested using the testing dataset. Two additional radiologists (readers 1 and 2) were involved in classifying radiological reports within the test dataset for testing purposes. The fine-tuned LLM, reader 1, and reader 2 demonstrated an accuracy of 0.979, 0.996, and 0.993, sensitivity for groups 0/1/2 of 0.988/0.947/0.943, 1.000/1.000/0.966, and 1.000/0.982/0.954, and time required for classification (s) of 105, 2312, and 3094 in under-sampled test dataset (n = 711), respectively. Fine-tuned LLM extracted patients with bone metastasis, demonstrating satisfactory performance that was comparable to or slightly lower than manual annotation by radiologists in a noticeably shorter time.

Backgrounds: Large language models (LLMs) are rapidly advancing and demonstrating high performance in understanding textual information, suggesting potential applications in interpreting patient histories and documented imaging findings. LLMs are advancing rapidly and an improvement in their diagnostic ability is expected. Furthermore, there has been a lack of comprehensive comparisons between LLMs from various manufacturers. Purpose: We tested the diagnostic performance of the latest three major LLMs (GPT-4o, Claude 3 Opus, and Gemini 1.5 Pro) using Radiology Diagnosis Please cases, a monthly diagnostic quiz series for radiology experts. Materials and Methods: Clinical history and imaging findings as provided textually by the case submitters were extracted from 324 quiz questions from Radiology Diagnosis Please cases. The GPT-4o, Claude 3 Opus, and Gemini 1.5 Pro generated the top three differential diagnoses. Diagnostic performance among the three LLMs were compared using Cochrane’s Q and post-hoc McNemar’s tests. Results: The diagnostic accuracies for the primary diagnosis were 41.0%, 54.0%, and 33.9% for GPT-4o, Claude 3 Opus, and Gemini 1.5 Pro, respectively. When considering the accuracy of any of the top three differential diagnoses, the rates improved to 49.4%, 62.0%, and 41.0%, respectively. Significant differences in diagnostic performance were observed among all pairs of the models. Conclusion: In a comparison of the latest LLMs, Claude 3 Opus outperformed GPT-4o and Gemini 1.5 Pro in solving radiology quiz cases. These models appear capable of assisting radiologists when supplied with accurate evaluations and worded descriptions of imaging findings by radiologists. Summary statement: Claude 3 Opus achieved the highest diagnostic accuracy, followed by GPT-4o and Gemini 1.5 Pro, in a comparison of their performance on 324 text-based Radiology Diagnosis Please cases.. Key Results: This study compared the diagnostic performances of the latest three major large language models, GPT-4o, Claude 3 Opus, and Gemini 1.5 Pro, using clinical history and textualized imaging findings in Radiology Diagnosis Please cases. The top three differential diagnoses generated by GPT-4o, Claude 3 Opus, and Gemini 1.5 Pro achieved diagnostic accuracies of 49.4%, 62.0%, and 41.0%, respectively, with statistically significant differences between each model’s performance.

Backgrounds Large language artificial intelligence models have showed its diagnostic performance based solely on textual information from clinical history and imaging findings. However, the extent of their performance when utilizing radiological images and providing differential diagnoses has yet to be investigated. Purpose We employed the latest version of Claude 3, Opus, released on March 4, 2024, to investigate its diagnostic performance in answering Radiology’s Diagnosis Please quiz questions under three conditions: (1) when provided with clinical history alone; (2) when given clinical history along with imaging findings; and (3) when supplied with clinical history and key images. Furthermore, we evaluated the diagnostic performance of the model when instructed to list differential diagnoses. Materials and Methods Claude 3 Opus was tasked with listing the primary diagnosis and two differential diagnoses for 322 quiz questions from Radiology’s “Diagnosis Please” cases, which included cases 1 to 322, published from 1998 to 2023. The analyses were carried out under the following input conditions: Condition 1: Submitter-provided clinical history (text) alone Condition 2: Submitter-provided clinical history and imaging findings (text) Condition 3: Submitter-provided clinical history (text) and key images (PDF files) We applied McNemar’s tests to evaluate differences in correct response rates for primary diagnoses across Conditions 1, 2, and 3. Results The correct primary diagnoses rates were 62/322 (19.3%), 178/322 (55.3%), and 93/322 (28.8%) for Conditions 1, 2, and 3, respectively. Additionally, Claude 3 Opus accurately provided the correct answer as a differential diagnosis in up to 22/322 (6.8%) of cases. There were statistically significant differences in correct response rates for primary diagnoses between all combinations of Conditions 1, 2, and 3 (pConclusion Claude 3 Opus demonstrated significantly improved diagnostic performance by inputting key images in addition to clinical history. The ability to list important differential diagnoses was also confirmed. Key Results This study investigated Claude 3 Opus’s performance in Radiology Diagnosis Please Cases using clinical history, key images, and imaging findings. Key images or imaging findings inputs significantly improved correct primary diagnoses from 19.3% to 28.8% or 55.5%, respectively. By having two additional differential diagnoses presented, total correct responses improved by 3.1–6.8%. Summary statement Large language AI model Claude 3 Opus demonstrated significantly improved diagnostic accuracy by adding key images with clinical history compared with clinical history alone.