Erickson, Bradley J.

Background and Aims: Among those with primary sclerosing cholangitis (PSC), perihilar cholangiocarcinoma (pCCA) is often diagnosed at a late stage and is a leading source of mortality. Detection of pCCA in PSC when curative action can be taken is challenging. Our aim was to create a deep learning model that analyzed MRI to detect early-stage pCCA and compare its diagnostic performance with expert radiologists. Approach and Results: We conducted a multicenter, international, retrospective cohort study involving adults with large duct PSC who underwent contrast-enhanced MRI. Senior abdominal radiologists reviewed the images. All patients with pCCA had early-stage cancer and were registered for liver transplantation. We trained a 3D DenseNet-121 model, a form of deep learning, using MRI images and assessed its performance in a separate test cohort. The study included 398 patients (training cohort n=150; test cohort n=248). pCCA was present in 230 individuals (training cohort n=64; test cohort n=166). In the test cohort, the respective performances of the model compared to the radiologists were: sensitivity 87.9% versus 50.0%, p <0.001; specificity 84.1% versus 100.0%, p <0.001; area under receiving operating curve 86.0% versus 75.0%, p <0.001. Even when a mass was absent, the model had a higher sensitivity for pCCA than radiologists (91.6% vs. 50.6%, p <0.001) and maintained good specificity (84.1%). Conclusions: The 3D DenseNet-121 MRI model effectively detects early-stage pCCA in PSC patients. Compared to expert radiologists, the model missed fewer cases of cancer.

Abstract Background Automated data extraction from echocardiography reports could facilitate large-scale registry creation and clinical surveillance of valvular heart diseases (VHD). We evaluated the performance of open-source large language models (LLMs) guided by prompt instructions and chain of thought (CoT) for this task. Methods From consecutive transthoracic echocardiographies performed in our center, we utilized 200 random reports from 2019 for prompt optimization and 1000 from 2023 for evaluation. Five instruction-tuned LLMs (Qwen2.0-72B, Llama3.0-70B, Mixtral8-46.7B, Llama3.0-8B, and Phi3.0-3.8B) were guided by prompt instructions with and without CoT to classify prosthetic valve presence and VHD severity. Performance was evaluated using classification metrics against expert-labeled ground truth. Mean squared error (MSE) was also calculated for predicted severity’s deviation from actual severity. Results With CoT prompting, Llama3.0-70B and Qwen2.0 achieved the highest performance (accuracy: 99.1% and 98.9% for VHD severity; 100% and 99.9% for prosthetic valve; MSE: 0.02 and 0.05, respectively). Smaller models showed lower accuracy for VHD severity (54.1%-85.9%) but maintained high accuracy for prosthetic valve detection (&gt;96%). Chain of thought reasoning yielded higher accuracy for larger models while increasing processing time from 2-25 to 67-154 seconds per report. Based on CoT reasonings, the wrong predictions were mainly due to model outputs being influenced by irrelevant information in the text or failure to follow the prompt instructions. Conclusions Our study demonstrates the near-perfect performance of open-source LLMs for automated echocardiography report interpretation with the purpose of registry formation and disease surveillance. While larger models achieved exceptional accuracy through prompt optimization, practical implementation requires balancing performance with computational efficiency.

To examine the association between post-stroke depression (PSD) and macrostructural and microstructural brain measures, and to explore whether changes in accelerometer-measured physical activity (PA) are associated with PSD, we conducted an exploratory study in UK Biobank with dementia-free participants diagnosed with at least one prior stroke. Eligible participants (n = 1186) completed an MRI scan. Depression was classified based on positive depression screening scores (PHQ-2 ≥ 3). Multivariate linear regression models assessed the relationships between depression and structural and diffusion measures generated from brain MRI scans. Logistic regression models were used to examine the relationship between accelerometer-measured daily PA and future depression (n = 367). Depression was positively associated with total white matter hyperintensities (WMHs) volume (standardized β [95% CI]—0.1339 [0.012, 0.256]; FDR-adjusted p-value—0.039), periventricular WMHs volume (standardized β [95% CI]—0.1351 [0.020, 0.250]; FDR-adjusted p-value—0.027), and reduced MD for commissural fibers (standardized β [95% CI]—−0.139 [−0.255, −0.024]; adjusted p-value—0.045). The odds of depression decreased by 0.3% for each daily minute spent in objectively measured light PA, while each minute spent in sleep from midnight to 6:00 AM was associated with a 0.9% decrease in the odds of depression. This early-stage analysis using a population cohort offers a scientific rationale for researchers using multimodal data sources to investigate the heterogenous nature of PSD and, potentially, identify stroke patients at risk of poor outcomes.

Abstract Deep learning is revolutionizing medical imaging analysis by enabling the classification of various pathoanatomical conditions at scale. Unfortunately, there have been a limited number of accurate and efficient machine learning (ML) algorithms that have been developed for the diagnostic workup of morphological hip pathologies, including developmental dysplasia of the hip and femoroacetabular impingement. The current study reports on the performance of a novel ML model with YOLOv5 and ConvNeXt-Tiny architecture in predicting the morphological features of these conditions, including cam deformity, ischial spine sign, dysplastic appearance, and other abnormalities. The model achieved 78.0% accuracy for detecting cam deformity, 87.2% for ischial spine sign, 76.6% for dysplasia, and 71.6% for all abnormalities combined. The model achieved an Area under the Receiver Operating Curve of 0.89 for ischial spine sign, 0.80 for cam deformity, 0.80 for dysplasia, and 0.81 for all abnormalities combined. Inter-rater agreement among surgeons, assessed using Gwet’s AC1, was substantial for dysplasia (0.83) and all abnormalities (0.88), and moderate for ischial spine sign (0.75) and cam deformity (0.61).

The Quantitative Imaging Biomarkers Alliance Profile for dynamic susceptibility contrast (DSC) MRI and its longitudinal claims for the precision of a DSC MRI–derived quantitative imaging biomarker to assess the relative cerebral blood volume changes in gliomas are summarized.

Abstract Aims The increasing need for diagnostic echocardiography tests presents challenges in preserving the quality and promptness of reports. While Large Language Models (LLMs) have proven effective in summarizing clinical texts, their application in echo remains underexplored. Methods and results Adult echocardiography studies, conducted at the Mayo Clinic from 1 January 2017 to 31 December 2017, were categorized into two groups: development (all Mayo locations except Arizona) and Arizona validation sets. We adapted open-source LLMs (Llama-2, MedAlpaca, Zephyr, and Flan-T5) using In-Context Learning and Quantized Low-Rank Adaptation fine-tuning (FT) for echo report summarization from ‘Findings’ to ‘Impressions.’ Against cardiologist-generated Impressions, the models’ performance was assessed both quantitatively with automatic metrics and qualitatively by cardiologists. The development dataset included 97 506 reports from 71 717 unique patients, predominantly male (55.4%), with an average age of 64.3 ± 15.8 years. EchoGPT, a fine-tuned Llama-2 model, outperformed other models with win rates ranging from 87% to 99% in various automatic metrics, and produced reports comparable to cardiologists in qualitative review (significantly preferred in conciseness (P &lt; 0.001), with no significant preference in completeness, correctness, and clinical utility). Correlations between automatic and human metrics were fair to modest, with the best being RadGraph F1 scores vs. clinical utility (r = 0.42) and automatic metrics showed insensitivity (0–5% drop) to changes in measurement numbers. Conclusion EchoGPT can generate draft reports for human review and approval, helping to streamline the workflow. However, scalable evaluation approaches dedicated to echo reports remains necessary.

An uncertainty-aware deep learning model developed using Mondrian conformal prediction demonstrated high performance in intracranial hemorrhage detection on noncontrast head CT volumes and high accuracy in identifying challenging cases.

Background: Although no tests are currently available to predict post-stroke depression (PSD), a risk factor for diminished quality of life and stroke-related death, emerging evidence suggests that ischemic stroke (IS) may induce behavioral changes by damaging connections throughout the brain rather than via localized injury. Hypothesis: We sought to determine if active (surveys) and passive (accelerometer, GPS, and other sensors) smartphone data could quantify exploratory behavioral proxies of PSD. Methods: Thirty-four adults with a recent IS (n = 12) or transient ischemic attack (TIA) (n = 22) diagnosis were followed for 3 months. Participants used the Beiwe app to complete a weekly Patient Health Questionnaire-8 (PHQ-8) survey and were evaluated every 30 days with the Montgomery-Asberg Depression Rating Scale (MADRS). Results: Preliminary analysis of app survey compliance was adequate (86%), and the quantity of data collected was greater for Android than iOS phones. PHQ-8 indicated depression (moderate or severe) in 20% of participants. MADRS highly correlated with the most recent PHQ-8 ( r = 0.92, 95% CI (0.19, 0.99), P = 0.028). Several exploratory passive physical activity and social measures showed statistically significant change over time and/or associations with PHQ-8 and MADRS. Conclusion: Low-burden smartphone data holds promise for developing novel IS and TIA outcome measures as well as identifying patients most at risk of PSD.