Quantitative cardiovascular PET/CT imaging is useful in the diagnosis of multiple cardiac perfusion and motion pathologies. The common approach for cardiac segmentation consists in using co-registered CT images, exploiting publicly available deep learning (DL)-based segmentation models. However, the mismatch between structural CT images and PET uptake limits the usefulness of these approaches. Besides, the performance of DL models is not consistent over low-dose or ultra-low-dose CT images commonly used in clinical PET/CT imaging. In this work, we developed a DL-based methodology to tackle this issue by segmenting directly cardiac PET images. This study included 406 cardiac PET images from 146 patients (43 ¹⁸F-FDG, 329 ¹³N-NH₃, and 37 ⁸²Rb images). Using previously trained DL nnU-Net models in our group, we segmented the whole heart and the three main cardiac components, namely the left myocardium (LM), left ventricle cavity (LV), and right ventricle (RV) on co-registered CT images. The segmentation was resampled to PET resolution and edited through a combination of automated image processing and manual correction. The corrected segmentation masks and SUV PET images were fed to a nnU-Net V2 pipeline to be trained in fivefold data split strategy by defining two tasks: task #1 for whole cardiac segmentation and task #2 for segmentation of three cardiac components. Fifteen cardiac images were used as external validation set. The DL delineated masks were compared with standard of reference masks using Dice coefficient, Jaccard distance, mean surface distance, and segment volume relative error (%). Task #1 average Dice coefficient in internal validation fivefold was 0.932 ± 0.033. The average Dice on the 15 external cases were comparable with the fivefold Dice reaching an average of 0.941 ± 0.018. Task #2 average Dice in fivefold validation was 0.88 ± 0.063, 0.828 ± 0.091, and 0.876 ± 0.062 for LM, LV, and RV, respectively. There was no statistically significant difference among the Dice coefficients, neither between images acquired by three radiotracers nor between the different folds (P-values >  > 0.05). The overall average volume prediction error in cardiac components segmentation was less than 2%. We developed an automated DL-based segmentation pipeline to segment the whole heart and cardiac components with acceptable accuracy and robust performance in the external test set and over three radiotracers used in nuclear cardiovascular imaging. The proposed methodology can overcome unreliable segmentations performed on CT images.

The early prediction of lymph node positivity (LN+) after neoadjuvant chemotherapy (NAC) is crucial for optimizing individualized treatment strategies. This study aimed to integrate radiomic features and clinical biomarkers through machine learning (ML) approaches to enhance prediction accuracy by focusing on patients with locally advanced gastric cancer (LAGC). We retrospectively enrolled 277 patients with LAGC and randomly divided them into training (n = 193) and validation (n = 84) sets at a 7:3 ratio. In total, 1,130 radiomics features were extracted from pre-treatment portal venous phase computed tomography scans. These features were linearly combined to develop a radiomics score (rad score) through feature engineering. Then, using the rad score and clinical biomarkers as input features, we applied simple statistical strategies (relying on a single ML model) and integrated statistical strategies (including classification model integration techniques, such as hard voting, soft voting, and stacking) to predict LN+ post-NAC. The diagnostic performance of the model was assessed using receiver operating characteristic curves with corresponding areas under the curve (AUC). Of all ML models, the stacking classifier, an integrated statistical strategy, exhibited the best performance, achieving an AUC of 0.859 for predicting LN+ in patients with LAGC. This predictive model was transformed into a publicly available online risk calculator. We developed a stacking classifier that integrates radiomics and clinical biomarkers to predict LN+ in patients with LAGC undergoing surgical resection, providing personalized treatment insights.

Accurate measurement of anterior segment parameters is crucial for diagnosing and managing ophthalmic conditions, such as glaucoma, cataracts, and refractive errors. However, traditional clinical measurement methods are often time-consuming, labor-intensive, and susceptible to inaccuracies. With the growing potential of artificial intelligence in ophthalmic diagnostics, this study aims to develop and evaluate a deep learning model capable of automatically extracting key points and precisely measuring multiple clinically significant anterior segment parameters from ultrasound biomicroscopy (UBM) images. These parameters include central corneal thickness (CCT), anterior chamber depth (ACD), pupil diameter (PD), angle-to-angle distance (ATA), sulcus-to-sulcus distance (STS), lens thickness (LT), and crystalline lens rise (CLR). A data set of 716 UBM anterior segment images was collected from Tianjin Medical University Eye Hospital. YOLOv8 was utilized to segment four key anatomical structures: cornea-sclera, anterior chamber, pupil, and iris-ciliary body-thereby enhancing the accuracy of keypoint localization. Only images with intact posterior capsule lentis were selected to create an effective data set for parameter measurement. Ten keypoints were localized across the data set, allowing the calculation of seven essential parameters. Control experiments were conducted to evaluate the impact of segmentation on measurement accuracy, with model predictions compared against clinical gold standards. The segmentation model achieved a mean IoU of 0.8836 and mPA of 0.9795. Following segmentation, the binary classification model attained an mAP of 0.9719, with a precision of 0.9260 and a recall of 0.9615. Keypoint localization exhibited a Euclidean distance error of 58.73 ± 63.04 μm, improving from the pre-segmentation error of 71.57 ± 67.36 μm. Localization mAP was 0.9826, with a precision of 0.9699, a recall of 0.9642 and an FPS of 32.64. In addition, parameter error analysis and Bland-Altman plots demonstrated improved agreement with clinical gold standards after segmentation. This deep learning approach for UBM image segmentation, keypoint localization, and parameter measurement is feasible, enhancing clinical diagnostic efficiency for anterior segment parameters.

Brain tumours continue to be a primary cause of worldwide death, highlighting the critical need for effective and accurate diagnostic tools. This article presents MK-YOLOv8, an innovative lightweight deep learning framework developed for the real-time detection and categorization of brain tumours from MRI images. Based on the YOLOv8 architecture, the proposed model incorporates Ghost Convolution, the C3Ghost module, and the SPPELAN module to improve feature extraction and substantially decrease computational complexity. An x-small object detection layer has been added, supporting precise detection of small and x-small tumours, which is crucial for early diagnosis. Trained on the Figshare Brain Tumour (FBT) dataset comprising (3,064) MRI images, MK-YOLOv8 achieved a mean Average Precision (mAP) of 99.1% at IoU (0.50) and 88.4% at IoU (0.50-0.95), outperforming YOLOv8 (98% and 78.8%, respectively). Glioma recall improved by 26%, underscoring the enhanced sensitivity to challenging tumour types. With a computational footprint of only 96.9 GFLOPs (representing 37.5% of YOYOLOv8x'sFLOPs) and utilizing 12.6 million parameters, a mere 18.5% of YOYOLOv8's parameters, MK-YOLOv8 delivers high efficiency with reduced resource demands. Also, it trained on the Br35H dataset (801 images) to guarantee the model's robustness and generalization; it achieved a mAP of 98.6% at IoU (0.50). The suggested model operates at 62 frames per second (FPS) and is suited for real-time clinical processes. These developments establish MK-YOLOv8 as an innovative framework, overcoming challenges in tiny tumour identification and providing a generalizable, adaptable, and precise detection approach for brain tumour diagnostics in clinical settings.

Cerebral aneurysm is a silent yet prevalent condition that affects a substantial portion of the global population. Aneurysms can develop due to various factors and present differently, necessitating diverse treatment approaches. Choosing the appropriate treatment upon diagnosis is paramount, as the severity of the disease dictates the course of action. The vulnerability of an aneurysm, particularly in the circle of Willis, is a critical concern; rupture can lead to irreversible consequences, including death. The primary objective of this study is to predict the rupture status of cerebral aneurysms using a comprehensive dataset that includes clinical, morphological, and hemodynamic data extracted from blood flow simulations of patients with actual vessels. Our goal is to provide valuable insights that can aid in treatment decision-making and potentially save the lives of future patients. Diagnosing and predicting the rupture status of aneurysms based solely on brain scans poses a significant challenge, often with limited accuracy, even for experienced physicians. However, harnessing statistical and machine learning (ML) techniques can enhance rupture prediction and treatment strategy selection. We employed a diverse set of supervised and unsupervised algorithms, training them on a database comprising over 700 cerebral aneurysms, which included 55 different parameters: 3 clinical, 35 morphological, and 17 hemodynamic features. Two of our models including stochastic gradient descent (SGD) and multi-layer perceptron (MLP) achieved a maximum area under the curve (AUC) of 0.86, a precision rate of 0.86, and a recall rate of 0.90 for prediction of cerebral aneurysm rupture. Given the sensitivity of the data and the critical nature of the condition, recall is a more vital parameter than accuracy and precision; our study achieved an acceptable recall score. Key features for rupture prediction included ellipticity index, low shear area ratio, and irregularity. Additionally, a one-dimensional CNN model predicted rupture status along a continuous spectrum, achieving 0.78 accuracy on the testing dataset, providing nuanced insights into rupture propensity.

We propose a multimodal spatiotemporal graph neural network (STG) framework to predict colorectal cancer liver metastasis (CRLM) progression. Current clinical models do not effectively integrate the tumor's spatial heterogeneity, dynamic evolution, and complex multimodal data relationships, limiting their predictive accuracy. Our STG framework combines preoperative CT imaging and clinical data into a heterogeneous graph structure, enabling joint modeling of tumor distribution and temporal evolution through spatial topology and cross-modal edges. The framework uses GraphSAGE to aggregate spatiotemporal neighborhood information and leverages supervised and contrastive learning strategies to enhance the model's ability to capture temporal features and improve robustness. A lightweight version of the model reduces parameter count by 78.55%, maintaining near-state-of-the-art performance. The model jointly optimizes recurrence risk regression and survival analysis tasks, with contrastive loss improving feature representational discriminability and cross-modal consistency. Experimental results on the MSKCC CRLM dataset show a time-adjacent accuracy of 85% and a mean absolute error of 1.1005, significantly outperforming existing methods. The innovative heterogeneous graph construction and spatiotemporal decoupling mechanism effectively uncover the associations between dynamic tumor microenvironment changes and prognosis, providing reliable quantitative support for personalized treatment decisions.

Tricuspid regurgitation jet velocity (TRJV) on echocardiography is used for screening patients with suspected pulmonary hypertension (PH). Artificial intelligence (AI) tools, such as the US2.AI, have been developed for automated evaluation of echocardiograms and can yield measurements that aid PH detection. This study evaluated the performance and utility of the US2.AI in a consecutive cohort of patients with suspected PH. 1031 patients who had been investigated for suspected PH between 2009-2021 were retrospectively identified from the ASPIRE registry. All patients had undergone echocardiography and right heart catheterisation (RHC). Based on RHC results, 771 (75%) patients with a mean pulmonary arterial pressure >20 mmHg were classified as having a diagnosis of PH (as per the 2022 European guidelines). Echocardiograms were evaluated manually and by the US2.AI tool to yield TRJV measurements. The AI tool demonstrated high interpretation yield, successfully measuring TRJV in 87% of echocardiograms. Manually and automatically derived TRJV values showed excellent agreement (intraclass correlation coefficient 0.94, 95% CI 0.94-0.95) with minimal bias (Bland-Altman analysis). Automated TRJV measurements showed equally high diagnostic accuracy for PH as manual measurements (area under the curve 0.88, 95% CI 0.84-0.90 <i>versus</i> 0.88, 95% CI 0.86-0.91). Automated TRJV measurements on echocardiography were similar to manual measurements, with similarly high and noninferior diagnostic accuracy for PH. These findings demonstrate that automated measurement of TRJV on echocardiography is feasible, accurate and reliable and support the implementation of AI-based approaches to echocardiogram evaluation and diagnostic imaging for PH.

Progression free survival (PFS) is a critical clinical outcome endpoint during cancer management and treatment evaluation. Yet, PFS is often missing from publicly available datasets due to the current subjective, expert, and time-intensive nature of generating PFS metrics. Given emerging research in multi-modal machine learning (ML), we explored the benefits and challenges associated with mining different electronic health record (EHR) data modalities and automating extraction of PFS metrics via ML algorithms. We analyzed EHR data from 92 pathology-proven GBM patients, obtaining 233 corticosteroid prescriptions, 2080 radiology reports, and 743 brain MRI scans. Three methods were developed to derive clinical PFS: 1) frequency analysis of corticosteroid prescriptions, 2) natural language processing (NLP) of reports, and 3) computer vision (CV) volumetric analysis of imaging. Outputs from these methods were compared to manually annotated clinical guideline PFS metrics. Employing data-driven methods, standalone progression rates were 63% (prescription), 78% (NLP), and 54% (CV), compared to the 99% progression rate from manually applied clinical guidelines using integrated data sources. The prescription method identified progression an average of 5.2 months later than the clinical standard, while the CV and NLP algorithms identified progression earlier by 2.6 and 6.9 months, respectively. While lesion growth is a clinical guideline progression indicator, only half of patients exhibited increasing contrast-enhancing tumor volumes during scan-based CV analysis. Our results indicate that data-driven algorithms can extract tumor progression outcomes from existing EHR data. However, ML methods are subject to varying availability bias, supporting contextual information, and pre-processing resource burdens that influence the extracted PFS endpoint distributions. Our scan-based CV results also suggest that the automation of clinical criteria may not align with human intuition. Our findings indicate a need for improved data source integration, validation, and revisiting of clinical criteria in parallel to multi-modal ML algorithm development.

It is unknown whether prediction models for lung function decline using computed tomography (CT) imaging-derived features from the upper lobes improve performance compared with globally derived features in individuals at risk of and with COPD. Individuals at risk (current or former smokers) and those with COPD from the Canadian Cohort Obstructive Lung Disease (CanCOLD) retrospective study, were investigated. A total of 103 CT features were extracted globally and regionally, and were used with 12 clinical features (demographics, questionnaires and spirometry) to predict rapid lung function decline for individuals at risk and those with COPD. Machine-learning models were evaluated in a hold-out test set using the area under the receiver operating characteristic curve (AUC) with DeLong's test for comparison. A total of 780 participants were included (n=276 at risk; n=298 Global Initiative for Chronic Obstructive Lung Disease (GOLD) 1 COPD; n=206 GOLD 2+ COPD). For predicting rapid lung function decline in those at risk, the upper-lobe CT model obtained a significantly higher AUC (AUC=0.80) than the lower-lobe CT model (AUC=0.63) and global model (AUC=0.66; p<0.05). For predicting rapid lung function decline in COPD, there was no significant differences between the upper-lobe (AUC=0.63), lower-lobe (AUC=0.59) or global CT features model (AUC=059; p>0.05). CT features extracted from the upper lobes obtained significantly improved prediction performance compared with globally extracted features for rapid lung function decline in early/mild COPD.

While several artificial intelligence (AI)-assisted medical imaging applications are reported in the recent orthopaedic literature, comparison of the clinical efficacy and utility of these applications is currently lacking. The aim of this systematic review is to evaluate the effectiveness and reliability of AI applications in orthopaedic imaging, focusing on their impact on diagnostic accuracy, image segmentation and operational efficiency across various imaging modalities. Based on the PRISMA guidelines, a comprehensive literature search of PubMed, Cochrane and Scopus databases was performed, using combinations of keywords and MeSH descriptors ('AI', 'ML', 'deep learning', 'orthopaedic surgery' and 'imaging') from inception to March 2024. Included were studies published between September 2018 and February 2024, which evaluated machine learning (ML) model effectiveness in improving orthopaedic imaging. Studies with insufficient data regarding the output variable used to assess the reliability of the ML model, those applying deterministic algorithms, unrelated topics, protocol studies, and other systematic reviews were excluded from the final synthesis. The Joanna Briggs Institute (JBI) Critical Appraisal tool and the Risk Of Bias In Non-randomised Studies-of Interventions (ROBINS-I) tool were applied for the assessment of bias among the included studies. The 53 included studies reported the use of 11.990.643 images from several diagnostic instruments. A total of 39 studies reported details in terms of the Dice Similarity Coefficient (DSC), while both accuracy and sensitivity were documented across 15 studies. Precision was reported by 14, specificity by nine, and the F1 score by four of the included studies. Three studies applied the area under the curve (AUC) method to evaluate ML model performance. Among the studies included in the final synthesis, Convolutional Neural Networks (CNN) emerged as the most frequently applied category of ML models, present in 17 studies (32%). The systematic review highlights the diverse application of AI in orthopaedic imaging, demonstrating the capability of various machine learning models in accurately segmenting and analysing orthopaedic images. The results indicate that AI models achieve high performance metrics across different imaging modalities. However, the current body of literature lacks comprehensive statistical analysis and randomized controlled trials, underscoring the need for further research to validate these findings in clinical settings. Systematic Review; Level of evidence IV.