J Exp Orthop. 2024 Dec 11;11(4):e70097. doi: 10.1002/jeo2.70097. eCollection 2024 Oct.

ABSTRACT

PURPOSE: The incidence of both primary and revision total knee arthroplasty (TKA) is expected to rise, making early recognition of TKA failure crucial to prevent extensive revision surgeries. This study aims to develop a deep learning (DL) model to predict TKA failure using radiographic images.

METHODS: Two patient cohorts who underwent primary TKA were retrospectively collected: one was used for the model development and the other for the external validation. Each cohort encompassed failed and non-failed subjects, according to the need for TKA revision surgery. Moreover, for each patient, one anteroposterior and one lateral radiographic view obtained during routine TKA follow-up, were considered. A transfer learning fine-tuning approach was employed. After pre-processing, the images were analyzed using a convolutional neuronal network (CNN) that was originally developed for predicting hip prosthesis failure and was based on the Densenet169 pre-trained on Imagenet. The model was tested on 20% of the images of the first cohort and externally validated on the images of the second cohort. Metrics, such as accuracy, sensitivity, specificity and area under the receiving operating characteristic curve (AUC), were calculated for the final assessment.

RESULTS: The trained model correctly classified 108 out of 127 images in the test set, providing a classification accuracy of 0.85, sensitivity of 0.80, specificity of 0.89 and AUC of 0.86. Moreover, the model correctly classified 1547 out of 1937 in the external validation set, providing a balanced accuracy of 0.79, sensitivity of 0.80, specificity of 0.78 and AUC of 0.86.

CONCLUSIONS: The present DL model predicts TKA failure with moderate accuracy, regardless of the cause of revision surgery. Additionally, the effectiveness of the transfer learning fine-tuning approach, leveraging a previously developed DL model for hip prosthesis failure, has been successfully demonstrated.

LEVEL OF EVIDENCE: Level III, diagnostic study.

PMID:39664926 | PMC:PMC11633713 | DOI:10.1002/jeo2.70097

Front Vet Sci. 2024 Nov 27;11:1453765. doi: 10.3389/fvets.2024.1453765. eCollection 2024.

ABSTRACT

Intervertebral disc disease is the most common spinal cord-related disease in dogs, caused by disc material protrusion or extrusion that compresses the spinal cord, leading to clinical symptoms. Diagnosis involves identifying radiographic signs such as intervertebral disc space narrowing, increased opacity of the intervertebral foramen, spondylosis deformans, and magnetic resonance imaging findings like spinal cord compression and lesions, alongside clinical symptoms and neurological examination findings. Intervertebral disc space narrowing on radiographs is the most common finding in intervertebral disc extrusion. This study aimed to develop a deep learning model to automatically recognize narrowed intervertebral disc space on caudal thoracic and lumbar X-ray images of dogs. In total, 241 caudal thoracic and lumbar lateral X-ray images from 142 dogs were used to develop and evaluate the model, which quantified intervertebral disc space distance and detected narrowing using a large-kernel one-dimensional convolutional neural network. When comparing veterinary clinicians and the deep learning model, the kappa value was 0.780, with 81.5% sensitivity and 95.6% specificity, showing substantial agreement. In conclusion, the deep learning model developed in this study, automatically and accurately quantified intervertebral disc space distance and detected narrowed sites in dogs, aiding in the initial screening of intervertebral disc disease and lesion localization.

PMID:39664893 | PMC:PMC11631885 | DOI:10.3389/fvets.2024.1453765

Transl Anim Sci. 2023 Mar 21;8:txad033. doi: 10.1093/tas/txad033. eCollection 2024.

ABSTRACT

This paper presents a visual deep learning approach to automatically determine hock and knee angles from sow images. Lameness is the second largest reason for culling of breeding herd females and relies on human observers to provide visual scoring for detection which can be slow, subjective, and inconsistent. A deep learning model classified and detected ten and two key body landmarks from the side and rear profile images, respectively (mean average precision = 0.94). Trigonometric-based formulae were derived to calculate hock and knee angles using the features extracted from the imagery. Automated angle measurements were compared with manual results from each image (average root mean square error [RMSE] = 4.13°), where all correlation slopes (average R 2 = 0.84) were statistically different from zero (P < 0.05); all automated measurements were in statistical agreement with manually collected measurements using the Bland-Altman procedure. This approach will be of interest to animal geneticists, scientists, and practitioners for obtaining objective angle measurements that can be factored into gilt replacement criteria to optimize sow breeding units.

PMID:39664862 | PMC:PMC11632189 | DOI:10.1093/tas/txad033

HeartRhythm Case Rep. 2024 Jul 25;10(10):757-760. doi: 10.1016/j.hrcr.2024.07.017. eCollection 2024 Oct.

NO ABSTRACT

PMID:39664847 | PMC:PMC11628775 | DOI:10.1016/j.hrcr.2024.07.017

Front Radiol. 2024 Nov 27;4:1495181. doi: 10.3389/fradi.2024.1495181. eCollection 2024.

ABSTRACT

INTRODUCTION: Task-based language fMRI is a non-invasive method of identifying brain regions subserving language that is used to plan neurosurgical resections which potentially encroach on eloquent regions. The use of unstructured fMRI paradigms, such as naturalistic fMRI, to map language is of increasing interest. Their analysis necessitates the use of alternative methods such as machine learning (ML) and deep learning (DL) because task regressors may be difficult to define in these paradigms.

METHODS: Using task-based language fMRI as a starting point, this study investigates the use of different categories of ML and DL algorithms to identify brain regions subserving language. Data comprising of seven task-based language fMRI paradigms were collected from 26 individuals, and ML and DL models were trained to classify voxel-wise fMRI time series.

RESULTS: The general machine learning and the interval-based methods were the most promising in identifying language areas using fMRI time series classification. The geneal machine learning method achieved a mean whole-brain Area Under the Receiver Operating Characteristic Curve (AUC) of 0.97 ± 0.03 , mean Dice coefficient of 0.6 ± 0.34 and mean Euclidean distance of 2.7 ± 2.4 mm between activation peaks across the evaluated regions of interest. The interval-based method achieved a mean whole-brain AUC of 0.96 ± 0.03 , mean Dice coefficient of 0.61 ± 0.33 and mean Euclidean distance of 3.3 ± 2.7 mm between activation peaks across the evaluated regions of interest.

DISCUSSION: This study demonstrates the utility of different ML and DL methods in classifying task-based language fMRI time series. A potential application of these methods is the identification of language activation from unstructured paradigms.

PMID:39664795 | PMC:PMC11631583 | DOI:10.3389/fradi.2024.1495181

Front Public Health. 2024 Nov 27;12:1443188. doi: 10.3389/fpubh.2024.1443188. eCollection 2024.

ABSTRACT

INTRODUCTION: Sarcopenia is a common age-related disease, defined as a decrease in muscle strength and function owing to reduced skeletal muscle. One way to diagnose sarcopenia is through gait analysis and foot-pressure imaging.

MOTIVATION AND RESEARCH GAP: We collected our own multimodal dataset from 100 subjects, consisting of both foot-pressure and skeleton data with real patients, which provides a unique resource for future studies aimed at more comprehensive analyses. While artificial intelligence has been employed for sarcopenia detection, previous studies have predominantly focused on skeleton-based datasets without exploring the combined potential of skeleton and foot pressure dataset. This study conducts separate experiments for foot-pressure and skeleton datasets, it demonstrates the potential of each data type in sarcopenia classification.

METHODS: This study had two components. First, we collected skeleton and foot-pressure datasets and classified them into sarcopenia and non-sarcopenia groups based on grip strength, gait performance, and appendicular skeletal muscle mass. Second, we performed experiments on the foot-pressure dataset using the ResNet-18 and spatiotemporal graph convolutional network (ST-GCN) models on the skeleton dataset to classify normal and abnormal gaits due to sarcopenia. For an accurate diagnosis, real-time walking of 100 participants was recorded at 30 fps as RGB + D images. The skeleton dataset was constructed by extracting 3D skeleton information comprising 25 feature points from the image, whereas the foot-pressure dataset was constructed by exerting pressure on the foot-pressure plates.

RESULTS: As a baseline evaluation, the accuracies of sarcopenia classification performance from foot-pressure image using Resnet-18 and skeleton sequences using ST-GCN were identified as 77.16 and 78.63%, respectively.

DISCUSSION: The experimental results demonstrated the potential applications of sarcopenia and non-sarcopenia classifications based on foot-pressure images and skeleton sequences.

PMID:39664552 | PMC:PMC11631742 | DOI:10.3389/fpubh.2024.1443188

Front Public Health. 2024 Nov 27;12:1413529. doi: 10.3389/fpubh.2024.1413529. eCollection 2024.

ABSTRACT

BACKGROUND: Screening for frequent hospitalizations in the community can help prevent super-utilizers from growing in the inpatient population. However, the determinants of frequent hospitalizations have not been systematically examined, their operational definitions have been inconsistent, and screening among community members lacks tools. Nor do we know if what determined frequent hospitalizations before COVID-19 continued to be the determinant of frequent hospitalizations at the height of the pandemic. Hence, the current study aims to identify determinants of frequent hospitalization and their screening items developed from the Comprehensive Geriatric Assessment (CGA), as our 273-item CGA is too lengthy to administer in full in community or primary care settings. The stability of the identified determinants will be examined in terms of the prospective validity of pre-COVID-selected items administered at the height of the pandemic.

METHODS: Comprehensive Geriatric Assessments (CGAs) were administered between 2016 and 2018 in the homes of 1,611 older adults aged 65+ years. Learning models were deployed to select CGA items to maximize the classification of different operational definitions of frequent hospitalizations, ranging from the most inclusive definition, wherein two or more hospitalizations over 2 years, to the most exclusive, wherein two or more hospitalizations must appear during year two, reflecting different care needs. In addition, the CGA items selected by the best-performing learning model were then developed into a random-forest-based scoring system for assessing frequent hospitalization risk, the validity of which was tested during 2018 and again prospectively between 2022 and 2023 in a sample of 329 older adults recruited from a district adjacent to where the CGAs were initially performed.

RESULTS: Seventeen items were selected from the CGA by our best-performing algorithm (DeepBoost), achieving 0.90 AUC in classifying operational definitions of frequent hospitalizations differing in temporal distributions and care needs. The number of medications prescribed and the need for assistance with emptying the bowel, housekeeping, transportation, and laundry were selected using the DeepBoost algorithm under the supervision of all operational definitions of frequent hospitalizations. On the other hand, reliance on walking aids, ability to balance on one's own, history of chronic obstructive pulmonary disease (COPD), and usage of social services were selected in the top 10 by all but the operational definitions that reflect the greatest care needs. The prospective validation of the original risk-scoring system using a sample recruited from a different district during the COVID-19 pandemic achieved an AUC of 0.82 in differentiating those rehospitalized twice or more over 2 years from those who were not.

CONCLUSION: A small subset of CGA items representing one's independence in aspects of (instrumental) activities of daily living, mobility, history of COPD, and social service utilization are sufficient for community members at risk of frequent hospitalization. The determinants of frequent hospitalization represented by the subset of CGA items remain relevant over the course of COVID-19 pandemic and across sociogeography.

PMID:39664532 | PMC:PMC11632619 | DOI:10.3389/fpubh.2024.1413529

Hellenic J Cardiol. 2024 Dec 9:S1109-9666(24)00261-6. doi: 10.1016/j.hjc.2024.12.002. Online ahead of print.

ABSTRACT

The integration of computational technologies into cardiology has significantly advanced the diagnosis and management of cardiovascular diseases. Computational cardiology, particularly through cardiovascular imaging and informatics, enables precise diagnosis of myocardial diseases by utilizing techniques such as echocardiography, cardiac magnetic resonance imaging, and computed tomography. Early-stage disease classification, especially in asymptomatic patients, benefits from these advancements, potentially altering disease progression and improving patient outcomes. Automatic segmentation of myocardial tissue using Deep Learning (DL) algorithms improves efficiency and consistency in analyzing large patient populations. Radiomic analysis can reveal subtle disease characteristics from medical images and can enhance disease detection, enable patient stratification, and facilitate monitoring of disease progression and treatment response. Radiomic biomarkers have already demonstrated high diagnostic accuracy in distinguishing myocardial pathologies and promise treatment individualization in cardiology, earlier disease detection, and disease monitoring. In this context, this narrative review explores the current state of the art in DL applications in medical imaging (CT, CMR, echocardiography and SPECT), focusing on automatic segmentation, radiomic feature phenotyping, and prediction of myocardial diseases, while also discussing challenges in integration of DL models in the clinical practice.

PMID:39662734 | DOI:10.1016/j.hjc.2024.12.002

BMC Cancer. 2024 Dec 11;24(1):1510. doi: 10.1186/s12885-024-13248-9.

ABSTRACT

BACKGROUND: In breast cancer, several gene expression assays have been developed to provide a more personalised treatment. This study focuses on the prediction of two molecular proliferation signatures: an 11-gene proliferation score and the MKI67 proliferation marker gene. The aim was to assess whether these could be predicted from digital whole slide images (WSIs) using deep learning models.

METHODS: WSIs and RNA-sequencing data from 819 invasive breast cancer patients were included for training, and models were evaluated on an internal test set of 172 cases as well as on 997 cases from a fully independent external test set. Two deep Convolutional Neural Network (CNN) models were optimised using WSIs and gene expression readouts from RNA-sequencing data of either the proliferation signature or the proliferation marker, and assessed using Spearman correlation (r). Prognostic performance was assessed through Cox proportional hazard modelling, estimating hazard ratios (HR).

RESULTS: Optimised CNNs successfully predicted the proliferation score and proliferation marker on the unseen internal test set (ρ = 0.691(p < 0.001) with R2 = 0.438, and ρ = 0.564 (p < 0.001) with R2 = 0.251 respectively) and on the external test set (ρ = 0.502 (p < 0.001) with R2 = 0.319, and ρ = 0.403 (p < 0.001) with R2 = 0.222 respectively). Patients with a high proliferation score or marker were significantly associated with a higher risk of recurrence or death in the external test set (HR = 1.65 (95% CI: 1.05-2.61) and HR = 1.84 (95% CI: 1.17-2.89), respectively).

CONCLUSIONS: The results from this study suggest that gene expression levels of proliferation scores can be predicted directly from breast cancer morphology in WSIs using CNNs and that the predictions provide prognostic information that could be used in research as well as in the clinical setting.

PMID:39663527 | DOI:10.1186/s12885-024-13248-9

Eye (Lond). 2024 Dec 11. doi: 10.1038/s41433-024-03532-0. Online ahead of print.

ABSTRACT

BACKGROUND: Investigate retinal fluid changes via a novel deep-learning algorithm in real-world patients receiving faricimab for the treatment of neovascular age-related macular degeneration (nAMD).

METHODS: Multicenter, retrospective chart review and optical coherence tomography (OCT) image upload from participating sites was conducted on patients treated with faricimab for nAMD from February 2022 to January 2024. The Notal OCT Analyzer (NOA) algorithm provided intraretinal, subretinal and total retinal fluid for each scan. Results were segregated based on treatment history and fluid compartments, allowing for multiple cross-sections of evaluation.

RESULTS: A total of 521 eyes were included at baseline. The previous treatments prior to faricimab were aflibercept, ranibizumab, bevacizumab, or treatment-naive for 52.3%, 21.0%, 13.3%, and 11.2% of the eyes, respectively. Of all 521 eyes, 49.9% demonstrated fluid reduction after one injection of faricimab. The mean fluid reduction after one injection was -60.7nL. The proportion of eyes that saw reduction in fluid compared to baseline after second, third, fourth and fifth faricimab injections were 54.4%, 51.9%, 51.4% and 52.2%, respectively. The mean (SD) retreatment interval after second, third, fourth and fifth faricimab injection were 53.4 (34.3), 56.6 (36.0), 57.1 (35.3) and 61.5 (40.2) days, respectively.

CONCLUSION: Deep-learning algorithms provide a novel tool for evaluating precise quantification of retinal fluid after treatment of nAMD with faricimab. Faricimab demonstrates reduction of retinal fluid in multiple groups after just one injection and sustains this response after multiple treatments, along with providing increases in treatment intervals between subsequent injections.

PMID:39663398 | DOI:10.1038/s41433-024-03532-0

Nat Cancer. 2024 Dec 11. doi: 10.1038/s43018-024-00869-z. Online ahead of print.

ABSTRACT

Large-scale omics profiling has uncovered a vast array of somatic mutations and cancer-associated proteins, posing substantial challenges for their functional interpretation. Here we present a network-based approach centered on FunMap, a pan-cancer functional network constructed using supervised machine learning on extensive proteomics and RNA sequencing data from 1,194 individuals spanning 11 cancer types. Comprising 10,525 protein-coding genes, FunMap connects functionally associated genes with unprecedented precision, surpassing traditional protein-protein interaction maps. Network analysis identifies functional protein modules, reveals a hierarchical structure linked to cancer hallmarks and clinical phenotypes, provides deeper insights into established cancer drivers and predicts functions for understudied cancer-associated proteins. Additionally, applying graph-neural-network-based deep learning to FunMap uncovers drivers with low mutation frequency. This study establishes FunMap as a powerful and unbiased tool for interpreting somatic mutations and understudied proteins, with broad implications for advancing cancer biology and informing therapeutic strategies.

PMID:39663389 | DOI:10.1038/s43018-024-00869-z

Pharm Res. 2024 Dec 12. doi: 10.1007/s11095-024-03800-4. Online ahead of print.

ABSTRACT

PURPOSE: Drug-induced liver injury, or DILI, affects numerous patients and also presents significant challenges in drug development. It has been attempted to predict DILI of a chemical by in silico approaches, including data-driven machine learning models. Herein, we report a recent DILI deep-learning effort that utilized our molecular representation concept by manifold embedding electronic attributes on a molecular surface.

METHODS: Local electronic attributes on a molecular surface were mapped to a lower-dimensional embedding of the surface manifold. Such an embedding was featurized in a matrix form and used in a deep-learning model as molecular input. The model was trained by a well-curated dataset and tested through cross-validations.

RESULTS: Our DILI prediction yielded superior results to the literature-reported efforts, suggesting that manifold embedding of electronic quantities on a molecular surface enables machine learning of molecular properties, including DILI.

CONCLUSIONS: The concept encodes the quantum information of a molecule that governs intermolecular interactions, potentially facilitating the deep-learning model development and training.

PMID:39663331 | DOI:10.1007/s11095-024-03800-4

J Imaging Inform Med. 2024 Dec 11. doi: 10.1007/s10278-024-01373-7. Online ahead of print.

ABSTRACT

Malnutrition is a commonly observed side effect in cancer patients, with a 30-85% worldwide prevalence in this population. Existing malnutrition screening tools miss ~ 20% of at-risk patients at initial screening and do not capture the abnormal body composition phenotype. Meanwhile, the gold-standard clinical criteria to diagnose malnutrition use changes in body composition as key parameters, particularly body fat and skeletal muscle mass loss. Diagnostic imaging, such as computed tomography (CT), is the gold-standard in analyzing body composition and typically accessible to cancer patients as part of the standard of care. In this study, we developed a deep learning-based body composition analysis approach over a diverse dataset of 200 abdominal/pelvic CT scans from cancer patients. The proposed approach segments adipose tissue and skeletal muscle using Swin UNEt TRansformers (Swin UNETR) at the third lumbar vertebrae (L3) level and automatically localizes L3 before segmentation. The proposed approach involves the first transformer-based deep learning model for body composition analysis and heatmap regression-based vertebra localization in cancer patients. Swin UNETR attained 0.92 Dice score in adipose tissue and 0.87 Dice score in skeletal muscle segmentation, significantly outperforming convolutional benchmarks including the 2D U-Net by 2-12% Dice score (p-values < 0.033). Moreover, Swin UNETR predictions showed high agreement with ground-truth areas of skeletal muscle and adipose tissue by 0.7-0.93 R2, highlighting its potential for accurate body composition analysis. We have presented an accurate body composition analysis based on CT imaging, which can enable the early detection of malnutrition in cancer patients and support timely interventions.

PMID:39663321 | DOI:10.1007/s10278-024-01373-7

J Imaging Inform Med. 2024 Dec 11. doi: 10.1007/s10278-024-01345-x. Online ahead of print.

ABSTRACT

Deep learning reconstruction (DLR) has exhibited potential in saving scan time. There is limited research on the evaluation of accelerated acquisition with DLR in staging rectal cancers. Our first objective was to explore the best DLR level in saving time through phantom experiments. Resolution and number of excitations (NEX) adjusted for different scan time, image quality of conventionally reconstructed T2W images were measured and compared with images reconstructed with different DLR level. The second objective was to explore the feasibility of accelerated T2W imaging with DLR in image quality and diagnostic performance for rectal cancer patients. 52 patients were prospectively enrolled to undergo accelerated acquisition reconstructed with highly-denoised DLR (DLR_H40sec) and conventional reconstruction (ConR2min). The image quality and diagnostic performance were evaluated by observers with varying experience and compared between protocols using κ statistics and area under the receiver operating characteristic curve (AUC). The phantom experiments demonstrated that DLR_H could achieve superior signal-to-noise ratio (SNR), detail conspicuity, sharpness, and less distortion within the least scan time. The DLR_H40sec images exhibited higher sharpness and SNR than ConR2min. The agreements with pathological TN-stages were improved using DLR_H40sec images compared to ConR2min (T: 0.846vs. 0.771, 0.825vs. 0.700, and 0.697vs. 0.512; N: 0.527vs. 0.521, 0.421vs. 0.348 and 0.517vs. 0.363 for junior, intermediate, and senior observes, respectively). Comparable AUCs to identify T3-4 and N1-2 tumors were achieved using DLR_H40sec and ConR2min images (P > 0.05). Consequently, with 2/3-time reduction, DLR_H40sec images showed improved image quality and comparable TN-staging performance to conventional T2W imaging for rectal cancer patients.

PMID:39663320 | DOI:10.1007/s10278-024-01345-x

J Imaging Inform Med. 2024 Dec 11. doi: 10.1007/s10278-024-01363-9. Online ahead of print.

ABSTRACT

Prostate cancer is one of the most prevalent male-specific diseases, where early and accurate diagnosis is essential for effective treatment and preventing disease progression. Assessing disease severity involves analyzing histological tissue samples, which are graded from 1 (healthy) to 5 (severely malignant) based on pathological features. However, traditional manual grading is labor-intensive and prone to variability. This study addresses the challenge of automating prostate cancer classification by proposing a novel histological grade analysis approach. The method integrates the gray-level co-occurrence matrix (GLCM) for extracting texture features with Haar wavelet modification to enhance feature quality. A convolutional neural network (CNN) is then employed for robust classification. The proposed method was evaluated using statistical and performance metrics, achieving an average accuracy of 97.3%, a precision of 98%, and an AUC of 0.95. These results underscore the effectiveness of the approach in accurately categorizing prostate tissue grades. This study demonstrates the potential of automated classification methods to support pathologists, enhance diagnostic precision, and improve clinical outcomes in prostate cancer care.

PMID:39663318 | DOI:10.1007/s10278-024-01363-9

J Imaging Inform Med. 2024 Dec 11. doi: 10.1007/s10278-024-01364-8. Online ahead of print.

ABSTRACT

Mammography images are widely used to detect non-palpable breast lesions or nodules, aiding in cancer prevention and enabling timely intervention when necessary. To support medical analysis, computer-aided detection systems can automate the segmentation of landmark structures, which is helpful in locating abnormalities and evaluating image acquisition adequacy. This paper presents a deep learning-based framework for segmenting the nipple, the pectoral muscle, the fibroglandular tissue, and the fatty tissue in standard-view mammography images. To the best of our knowledge, we introduce the largest dataset dedicated to mammography segmentation of key anatomical structures, specifically designed to train deep learning models for this task. Through comprehensive experiments, we evaluated various deep learning model architectures and training configurations, demonstrating robust segmentation performance across diverse and challenging cases. These results underscore the framework's potential for clinical integration. In our experiments, four semantic segmentation architectures were compared, all showing suitability for the target problem, thereby offering flexibility in model selection. Beyond segmentation, we introduce a suite of applications derived from this framework to assist in clinical assessments. These include automating tasks such as multi-view lesion registration and anatomical position estimation, evaluating image acquisition quality, measuring breast density, and enhancing visualization of breast tissues, thus addressing critical needs in breast cancer screening and diagnosis.

PMID:39663317 | DOI:10.1007/s10278-024-01364-8

J Imaging Inform Med. 2024 Dec 11. doi: 10.1007/s10278-024-01358-6. Online ahead of print.

ABSTRACT

Ultrasound imaging is the most cost-effective approach for the early detection of rectal cancer, which is a high-risk cancer. Our goal was to design an effective method that can accurately identify and segment rectal tumours in ultrasound images, thereby facilitating rectal cancer diagnoses for physicians. This would allow physicians to devote more time to determining whether the tumour is benign or malignant and whether it has metastasized rather than merely confirming its presence. Data originated from the Sichuan Province Cancer Hospital. The test, training, and validation sets were composed of 53 patients with 173 images, 195 patients with 1247 images, and 20 patients with 87 images, respectively. We created a deep learning network architecture consisting of encoders and decoders. To enhance global information capture, we substituted traditional convolutional decoders with global attention decoders and incorporated effective channel information fusion for multiscale information integration. The Dice coefficient (DSC) of the proposed model was 75.49%, which was 4.03% greater than that of the benchmark model, and the Hausdorff distance 95(HD95) was 24.75, which was 8.43 lower than that of the benchmark model. The paired t-test statistically confirmed the significance of the difference between our model and the benchmark model, with a p-value less than 0.05. The proposed method effectively identifies and segments rectal tumours of diverse shapes. Furthermore, it distinguishes between normal rectal images and those containing tumours. Therefore, after consultation with physicians, we believe that our method can effectively assist physicians in diagnosing rectal tumours via ultrasound.

PMID:39663316 | DOI:10.1007/s10278-024-01358-6

BMJ Open Ophthalmol. 2024 Dec 11;9(1):e001873. doi: 10.1136/bmjophth-2024-001873.

ABSTRACT

AIM: To evaluate the efficacy of automated machine learning (AutoML) models in detecting fundus diseases using ocular B-scan ultrasound images.

METHODS: Ophthalmologists annotated two B-scan ultrasound image datasets to develop three AutoML models-single-label, multi-class single-label and multi-label-on the Vertex artificial intelligence (AI) platform. Performance of these models was compared among themselves and against existing bespoke models for binary classification tasks.

RESULTS: The training set involved 3938 images from 1378 patients, while batch predictions used an additional set of 336 images from 180 patients. The single-label AutoML model, trained on normal and abnormal fundus images, achieved an area under the precision-recall curve (AUPRC) of 0.9943. The multi-class single-label model, focused on single-pathology images, recorded an AUPRC of 0.9617, with performance metrics of these two single-label models proving comparable to those of previously published models. The multi-label model, designed to detect both single and multiple pathologies, posted an AUPRC of 0.9650. Pathology classification AUPRCs for the multi-class single-label model ranged from 0.9277 to 1.0000 and from 0.8780 to 0.9980 for the multi-label model. Batch prediction accuracies ranged from 86.57% to 97.65% for various fundus conditions in the multi-label AutoML model. Statistical analysis demonstrated that the single-label model significantly outperformed the other two models in all evaluated metrics (p<0.05).

CONCLUSION: AutoML models, developed by clinicians, effectively detected multiple fundus lesions with performance on par with that of deep-learning models crafted by AI specialists. This underscores AutoML's potential to revolutionise ophthalmologic diagnostics, facilitating broader accessibility and application of sophisticated diagnostic technologies.

PMID:39663141 | DOI:10.1136/bmjophth-2024-001873

J Med Internet Res. 2024 Dec 11;26:e63892. doi: 10.2196/63892.

ABSTRACT

BACKGROUND: Hospital call centers play a critical role in providing support and information to patients with cancer, making it crucial to effectively identify and understand patient intent during consultations. However, operational efficiency and standardization of telephone consultations, particularly when categorizing diverse patient inquiries, remain significant challenges. While traditional deep learning models like long short-term memory (LSTM) and bidirectional encoder representations from transformers (BERT) have been used to address these issues, they heavily depend on annotated datasets, which are labor-intensive and time-consuming to generate. Large language models (LLMs) like GPT-4, with their in-context learning capabilities, offer a promising alternative for classifying patient intent without requiring extensive retraining.

OBJECTIVE: This study evaluates the performance of GPT-4 in classifying the purpose of telephone consultations of patients with cancer. In addition, it compares the performance of GPT-4 to that of discriminative models, such as LSTM and BERT, with a particular focus on their ability to manage ambiguous and complex queries.

METHODS: We used a dataset of 430,355 sentences from telephone consultations with patients with cancer between 2016 and 2020. LSTM and BERT models were trained on 300,000 sentences using supervised learning, while GPT-4 was applied using zero-shot and few-shot approaches without explicit retraining. The accuracy of each model was compared using 1,000 randomly selected sentences from 2020 onward, with special attention paid to how each model handled ambiguous or uncertain queries.

RESULTS: GPT-4, which uses only a few examples (a few shots), attained a remarkable accuracy of 85.2%, considerably outperforming the LSTM and BERT models, which achieved accuracies of 73.7% and 71.3%, respectively. Notably, categories such as "Treatment," "Rescheduling," and "Symptoms" involve multiple contexts and exhibit significant complexity. GPT-4 demonstrated more than 15% superior performance in handling ambiguous queries in these categories. In addition, GPT-4 excelled in categories like "Records" and "Routine," where contextual clues were clear, outperforming the discriminative models. These findings emphasize the potential of LLMs, particularly GPT-4, for interpreting complicated patient interactions during cancer-related telephone consultations.

CONCLUSIONS: This study shows the potential of GPT-4 to significantly improve the classification of patient intent in cancer-related telephone oncological consultations. GPT-4's ability to handle complex and ambiguous queries without extensive retraining provides a substantial advantage over discriminative models like LSTM and BERT. While GPT-4 demonstrates strong performance in various areas, further refinement of prompt design and category definitions is necessary to fully leverage its capabilities in practical health care applications. Future research will explore the integration of LLMs like GPT-4 into hybrid systems that combine human oversight with artificial intelligence-driven technologies.

PMID:39661975 | DOI:10.2196/63892

J Med Internet Res. 2024 Dec 11;26:e55986. doi: 10.2196/55986.

ABSTRACT

BACKGROUND: Real-time monitoring of pediatric epileptic seizures poses a significant challenge in clinical practice. In recent years, machine learning (ML) has attracted substantial attention from researchers for diagnosing and treating neurological diseases, leading to its application for detecting pediatric epileptic seizures. However, systematic evidence substantiating its feasibility remains limited.

OBJECTIVE: This systematic review aimed to consolidate the existing evidence regarding the effectiveness of ML in monitoring pediatric epileptic seizures with an effort to provide an evidence-based foundation for the development and enhancement of intelligent tools in the future.

METHODS: We conducted a systematic search of the PubMed, Cochrane, Embase, and Web of Science databases for original studies focused on the detection of pediatric epileptic seizures using ML, with a cutoff date of August 27, 2023. The risk of bias in eligible studies was assessed using the QUADAS-2 (Quality Assessment of Diagnostic Accuracy Studies-2). Meta-analyses were performed to evaluate the C-index and the diagnostic 4-grid table, using a bivariate mixed-effects model for the latter. We also examined publication bias for the C-index by using funnel plots and the Egger test.

RESULTS: This systematic review included 28 original studies, with 15 studies on ML and 13 on deep learning (DL). All these models were based on electroencephalography data of children. The pooled C-index, sensitivity, specificity, and accuracy of ML in the training set were 0.76 (95% CI 0.69-0.82), 0.77 (95% CI 0.73-0.80), 0.74 (95% CI 0.70-0.77), and 0.75 (95% CI 0.72-0.77), respectively. In the validation set, the pooled C-index, sensitivity, specificity, and accuracy of ML were 0.73 (95% CI 0.67-0.79), 0.88 (95% CI 0.83-0.91), 0.83 (95% CI 0.71-0.90), and 0.78 (95% CI 0.73-0.82), respectively. Meanwhile, the pooled C-index of DL in the validation set was 0.91 (95% CI 0.88-0.94), with sensitivity, specificity, and accuracy being 0.89 (95% CI 0.85-0.91), 0.91 (95% CI 0.88-0.93), and 0.89 (95% CI 0.86-0.92), respectively.

CONCLUSIONS: Our systematic review demonstrates promising accuracy of artificial intelligence methods in epilepsy detection. DL appears to offer higher detection accuracy than ML. These findings support the development of DL-based early-warning tools in future research.

TRIAL REGISTRATION: PROSPERO CRD42023467260; https://www.crd.york.ac.uk/prospero/display_record.php?ID=CRD42023467260.

PMID:39661965 | DOI:10.2196/55986