Deep learning models require large amounts of annotated data, which are hard to obtain in the medical field, as the annotation process is laborious and depends on expert knowledge. This data scarcity hinders a model's ability to generalise effectively on unseen data, and recently, foundation models pretrained on large datasets have been proposed as a promising solution. RadiologyNET is a custom medical dataset that comprises 1,902,414 medical images covering various body parts and modalities of image acquisition. We used the RadiologyNET dataset to pretrain several popular architectures (ResNet18, ResNet34, ResNet50, VGG16, EfficientNetB3, EfficientNetB4, InceptionV3, DenseNet121, MobileNetV3Small and MobileNetV3Large). We compared the performance of ImageNet and RadiologyNET foundation models against training from randomly initialiased weights on several publicly available medical datasets: (i) Segmentation-LUng Nodule Analysis Challenge, (ii) Regression-RSNA Pediatric Bone Age Challenge, (iii) Binary classification-GRAZPEDWRI-DX and COVID-19 datasets, and (iv) Multiclass classification-Brain Tumor MRI dataset. Our results indicate that RadiologyNET-pretrained models generally perform similarly to ImageNet models, with some advantages in resource-limited settings. However, ImageNet-pretrained models showed competitive performance when fine-tuned on sufficient data. The impact of modality diversity on model performance was tested, with the results varying across tasks, highlighting the importance of aligning pretraining data with downstream applications. Based on our findings, we provide guidelines for using foundation models in medical applications and publicly release our RadiologyNET-pretrained models to support further research and development in the field. The models are available at https://github.com/AIlab-RITEH/RadiologyNET-TL-models .

Anterior Cruciate Ligament (ACL) tears are common in sports and can provide noteworthy health issues. Therefore, accurately diagnosing of tears is important for the early and proper treatment. However, traditional diagnostic methods, such as clinical assessments and MRI, have limitations in terms of accuracy and efficiency. This study introduces a new diagnostic approach by combining of the deep learning architecture Res2Net with an improved version of the Lévy flight distribution (ILFD) to improve the detection of ACL tears in knee MRI images. The Res2Net model is known for its ability to extract important features and classify them effectively. By optimizing the model using the ILFD algorithm, the diagnostic efficiency is greatly improved. For validation of the proposed model's efficiency, it has been applied into two standard datasets including Stanford University Medical Center and Clinical Hospital Centre Rijeka. Comparative analysis with existing diagnostic methods, including 14 layers ResNet-14, Compact Parallel Deep Convolutional Neural Network (CPDCNN), Convolutional Neural Network (CNN), Generative Adversarial Network (GAN), and combined CNN and Modified Golden Search Algorithm (CNN/MGSA) shows that the suggested Res2Net/ILFD model performs better in various metrics, including precision, recall, accuracy, f1-score, and specificity, and Matthews correlation coefficient.

This study aimed to develop and validate convolutional neural network (CNN) models for distinguishing follicular thyroid carcinoma (FTC) from follicular thyroid adenoma (FTA). Additionally, this current study compared the performance of CNN models with the American College of Radiology Thyroid Imaging Reporting and Data System (ACR-TIRADS) and Chinese Thyroid Imaging Reporting and Data System (C-TIRADS) ultrasound-based malignancy risk stratification systems. A total of 327 eligible patients with FTC and FTA who underwent preoperative thyroid ultrasound examination were retrospectively enrolled between August 2017, and August 2024. Patients were randomly assigned to a training cohort (n = 263) and a test cohort (n = 64) in an 8:2 ratio using stratified sampling. Five CNN models, including VGG16, ResNet101, MobileNetV2, ResNet152, and ResNet50, pre-trained with ImageNet, were developed and tested to distinguish FTC from FTA. The CNN models exhibited good performance, yielding areas under the receiver operating characteristic curve (AUC) ranging from 0.64 to 0.77. The ResNet152 model demonstrated the highest AUC (0.77; 95% CI, 0.67-0.87) for distinguishing between FTC and FTA. Decision curve and calibration curve analyses demonstrated the models' favorable clinical value and calibration. Furthermore, when comparing the performance of the developed models with that of the C-TIRADS and ACR-TIRADS systems, the models developed in this study demonstrated superior performance. This can potentially guide appropriate management of FTC in patients with follicular neoplasms.

Introduction Contrast enema (CE) is widely used in the evaluation of suspected Hirschsprung disease (HD). Deep learning is a promising tool to standardize image assessment and support clinical decision-making. This study assesses the diagnostic performance of a deep neural network (DNN), with and without clinical data, and compares its interpretation with that of pediatric surgeons and radiologists. Materials and Methods In this retrospective study, 1471 contrast enema images from patients <15 years were analysed, with 218 images used for testing. A deep neural network, pediatric radiologists, and surgeons independently reviewed the testing set, with and without clinical data. Diagnostic performance was assessed using ROC and PR curves, and interobserver agreement was evaluated using Fleiss' kappa. Results The deep neural network achieved high diagnostic accuracy (AUC-ROC = 0.87) in contrast enema interpretation, with improved performance when combining anteroposterior and lateral images (AUC-ROC = 0.92). Clinical data integration further enhanced model sensitivity and negative predictive value. The super-surgeon (majority voting of colorectal surgeons) outperformed most individual clinicians (sensitivity 81.8%, specificity 79.1%), while the super-radiologist (majority voting of radiologist) showed moderate accuracy. Interobserver analysis revealed strong agreement between the model and surgeons (Cohen's kappa = 0.73), and overall consistency among experts and the model (Fleiss' kappa = 0.62). Conclusions AI-assisted CE interpretation achieved higher specificity and comparable sensitivity to those of the clinicians. Its consistent performance and substantial agreement with experts support its potential role in improving CE assessment in HD.

BackgroundThe precise pneumoconiosis staging suffers from progressive pair label noise (PPLN) in chest X-ray datasets, because adjacent stages are confused due to unidentifialble and diffuse opacities in the lung fields. As deep neural networks are employed to aid the disease staging, the performance is degraded under such label noise.ObjectiveThis study improves the effectiveness of pneumoconiosis staging by mitigating the impact of PPLN through network architecture refinement and sample selection mechanism adjustment.MethodsWe propose a novel multi-branch architecture that incorporates the dual-threshold sample selection. Several auxiliary branches are integrated in a two-phase module to learn and predict the <i>progressive feature tendency</i>. A novel difference-based metric is introduced to iteratively obtained the instance-specific thresholds as a complementary criterion of dynamic sample selection. All the samples are finally partitioned into <i>clean</i> and <i>hard</i> sets according to dual-threshold criteria and treated differently by loss functions with penalty terms.ResultsCompared with the state-of-the-art, the proposed method obtains the best metrics (accuracy: 90.92%, precision: 84.25%, sensitivity: 81.11%, F1-score: 82.06%, and AUC: 94.64%) under real-world PPLN, and is less sensitive to the rise of synthetic PPLN rate. An ablation study validates the respective contributions of critical modules and demonstrates how variations of essential hyperparameters affect model performance.ConclusionsThe proposed method achieves substantial effectiveness and robustness against PPLN in pneumoconiosis dataset, and can further assist physicians in diagnosing the disease with a higher accuracy and confidence.

In this study, the authors aimed to establish a supervised machine learning (ML) model based on multiple tree-based algorithms to predict long-term biochemical outcomes and intervention-free survival (IFS) after endonasal transsphenoidal surgery (ETS) in patients with Cushing's disease (CD). The medical records of patients who underwent ETS for CD between 2013 and 2023 were reviewed. Data were collected on the patient's baseline characteristics, intervention details, histopathology, surgical outcomes, and postoperative endocrine functions. The study's primary outcome was IFS, and the therapeutic outcomes were labeled as "under control" or "treatment failure," depending on whether additional therapeutic interventions after primary ETS were required. The decision tree and random forest classifiers were trained and tested to predict long-term IFS based on unseen data, using an 80/20 cohort split. Data from 150 patients, with a median follow-up period of 56 months, were extracted. In the cohort, 42 (28%) patients required additional intervention for persistent or recurrent CD. Consequently, the IFS rates following ETS alone were 83% at 3 years and 78% at 5 years. Multivariable Cox proportional hazards analysis demonstrated that a smaller tumor diameter that could be detected by MRI (hazard ratio 0.95, 95% CI 0.90-0.99; p = 0.047) was significantly associated with greater IFS. However, the lack of tumor detection on MRI was a poor predictor. The ML-based model using a decision tree model displayed 91% accuracy (95% CI 0.70-0.94, sensitivity 87.0%, specificity 89.0%) in predicting IFS in the unseen test dataset. Random forest analysis revealed that tumor size (mean minimal depth 1.67), Knosp grade (1.75), patient age (1.80), and BMI (1.99) were the four most significant predictors of long-term IFS. The ML algorithm could predict long-term postoperative endocrinological remission in CD with high accuracy, indicating that prognosis may vary not only with previously reported factors such as tumor size, Knosp grade, gross-total resection, and patient age but also with BMI. The decision tree flowchart could potentially stratify patients with CD before ETS, allowing for the selection of personalized treatment options and thereby assisting in determining treatment plans for these patients. This ML model may lead to a deeper understanding of the complex mechanisms of CD by uncovering patterns embedded within the data.

Cardiac amyloidosis (CA) is a severe condition characterized by amyloid fibril deposition in the myocardium, leading to restrictive cardiomyopathy and heart failure. Differentiating between amyloidosis subtypes is crucial due to distinct treatment strategies. The individual conventional diagnostic methods lack the accuracy needed for effective subtype identification. This study aimed to evaluate the efficacy of combining ¹¹C-PiB PET/CT and ^99mTc-DPD scintigraphy in detecting CA and distinguishing between its main subtypes, light chain (AL) and transthyretin (ATTR) amyloidosis while assessing the association of imaging findings with patient prognosis. We retrospectively evaluated the diagnostic efficacy of combining ¹¹C-PiB PET/CT and ^99mTc-DPD scintigraphy in a cohort of 50 patients with clinical suspicion of CA. Semi-quantitative imaging markers were extracted from the images. Diagnostic performance was calculated against biopsy results or genetic testing. Both machine learning models and a rationale-based model were developed to detect CA and classify subtypes. Survival prediction over five years was assessed using a random survival forest model. Prognostic value was assessed using Kaplan-Meier estimators and Cox proportional hazards models. The combined imaging approach significantly improved diagnostic accuracy, with ¹¹C-PiB PET and ^99mTc-DPD scintigraphy showing complementary strengths in detecting AL and ATTR, respectively. The machine learning model achieved an AUC of 0.94 (95% CI 0.93-0.95) for CA subtype differentiation, while the rationale-based model demonstrated strong diagnostic ability with AUCs of 0.95 (95% CI 0.88-1.00) for ATTR and 0.88 (95% CI 0.770-0.961) for AL. Survival prediction models identified key prognostic markers, with significant stratification of overall mortality based on predicted survival (p value = 0.006; adj HR 2.43 [95% CI 1.03-5.71]). The integration of ¹¹C-PiB PET/CT and ^99mTc-DPD scintigraphy, supported by both machine learning and rationale-based models, enhances the diagnostic accuracy and prognostic assessment of cardiac amyloidosis, with significant implications for clinical practice.

Deep learning-based diagnostic models often suffer performance drops due to distribution shifts between training (source) and test (target) domains. Collecting and labeling sufficient target domain data for model retraining represents an optimal solution, yet is limited by time and scarce resources. Active learning (AL) offers an efficient approach to reduce annotation costs while maintaining performance, but struggles to handle the challenge posed by distribution variations across different datasets. In this study, we propose a novel unsupervised Active learning framework for Domain Adaptation, named ADAptation, which efficiently selects informative samples from multi-domain data pools under limited annotation budget. As a fundamental step, our method first utilizes the distribution homogenization capabilities of diffusion models to bridge cross-dataset gaps by translating target images into source-domain style. We then introduce two key innovations: (a) a hypersphere-constrained contrastive learning network for compact feature clustering, and (b) a dual-scoring mechanism that quantifies and balances sample uncertainty and representativeness. Extensive experiments on four breast ultrasound datasets (three public and one in-house/multi-center) across five common deep classifiers demonstrate that our method surpasses existing strong AL-based competitors, validating its effectiveness and generalization for clinical domain adaptation. The code is available at the anonymized link: https://github.com/miccai25-966/ADAptation.

Major Depressive Disorder (MDD) diagnosis mainly relies on subjective self-reporting and clinical assessments. Resting-state functional magnetic resonance imaging (rs-fMRI) and its analysis of Effective Connectivity (EC) offer a quantitative approach to understand the directional interactions between brain regions, presenting a potential objective method for MDD classification. Granger causality analysis was used to extract EC features from a large, multi-site rs-fMRI dataset of MDD patients. The ComBat algorithm was applied to adjust for site differences, while multivariate linear regression was employed to control for age and sex differences. Discriminative EC features for MDD were identified using two-sample t-tests and model-based feature selection, with the LightGBM algorithm being used for classification. The performance and generalizability of the model was evaluated using nested five-fold cross-validation and tested for generalizability on an independent dataset. Ninety-seven EC features belonging to the cerebellum and front-temporal regions were identified as highly discriminative for MDD. The classification model using these features achieved an accuracy of 94.35 %, with a sensitivity of 93.52 % and specificity of 95.25 % in cross-validation. Generalization of the model to an independent dataset resulted in an accuracy of 94.74 %, sensitivity of 90.59 %, and specificity of 96.75 %. The study demonstrates that EC features from rs-fMRI can effectively discriminate MDD from healthy controls, suggesting that EC analysis could be a valuable tool in assisting the clinical diagnosis of MDD. This method shows promise in enhancing the objectivity of MDD diagnosis through the use of neuroimaging biomarkers.

Nonhuman primates, particularly rhesus macaques, have served as crucial animal models for investigating complex brain functions. While previous studies have explored neural activity features in macaques, the gyro-sulcal functional dissociation characteristics are largely unknown. In this study, we employ a deep learning model named one-dimensional convolutional neural network to differentiate resting state functional magnetic resonance imaging signals between gyri and sulci in macaque brains, and further investigate the frequency-specific dissociations between gyri and sulci inferred from the power spectral density of resting state functional magnetic resonance imaging. Experimental results based on a large cohort of 440 macaques from two independent sites demonstrate substantial frequency-specific dissociation between gyral and sulcal signals at both whole-brain and regional levels. The magnitude of gyral power spectral density is significantly larger than that of sulcal power spectral density within the range of 0.01 to 0.1 Hz, suggesting that gyri and sulci may play distinct roles as the global hubs and local processing units for functional activity transmission and interaction in macaque brains. In conclusion, our study has established one of the first power spectrum maps of gyro-sulcal functional activity dissociation in macaque brains, providing a novel perspective for systematically exploring the neural mechanism of functional dissociation in mammalian brains.