The development of larger models for medical image analysis has led to increased performance. However, it also affected our ability to explain and validate model decisions. Models can use non-relevant parts of images, also called spurious correlations or shortcuts, to obtain high performance on benchmark datasets but fail in real-world scenarios. In this work, we challenge the capacity of convolutional neural networks (CNN) to classify chest X-rays and eye fundus images while masking out clinically relevant parts of the image. We show that all models trained on the PadChest dataset, irrespective of the masking strategy, are able to obtain an area under the curve (AUC) above random. Moreover, the models trained on full images obtain good performance on images without the region of interest (ROI), even superior to the one obtained on images only containing the ROI. We also reveal a possible spurious correlation in the Chákṣu dataset while the performances are more aligned with the expectation of an unbiased model. We go beyond the performance analysis with the usage of the explainability method SHAP and the analysis of embeddings. We asked a radiology resident to interpret chest X-rays under different masking to complement our findings with clinical knowledge.

The aim of this study was to develop a hybrid diagnostic framework integrating radiomic and deep features for accurate and reproducible detection and classification of wrist fractures using X-ray images. A total of 3,537 X-ray images, including 1,871 fracture and 1,666 non-fracture cases, were collected from three healthcare centers. Radiomic features were extracted using the PyRadiomics library, and deep features were derived from the bottleneck layer of an autoencoder. Both feature modalities underwent reliability assessment via Intraclass Correlation Coefficient (ICC) and cosine similarity. Feature selection methods, including ANOVA, Mutual Information (MI), Principal Component Analysis (PCA), and Recursive Feature Elimination (RFE), were applied to optimize the feature set. Classifiers such as XGBoost, CatBoost, Random Forest, and a Voting Classifier were used to evaluate diagnostic performance. The dataset was divided into training (70%) and testing (30%) sets, and metrics such as accuracy, sensitivity, and AUC-ROC were used for evaluation. The combined radiomic and deep feature approach consistently outperformed standalone methods. The Voting Classifier paired with MI achieved the highest performance, with a test accuracy of 95%, sensitivity of 94%, and AUC-ROC of 96%. The end-to-end model achieved competitive results with an accuracy of 93% and AUC-ROC of 94%. SHAP analysis and t-SNE visualizations confirmed the interpretability and robustness of the selected features. This hybrid framework demonstrates the potential for integrating radiomic and deep features to enhance diagnostic performance for wrist and forearm fractures, providing a reliable and interpretable solution suitable for clinical applications.

The International Society of Ultrasound advocates Intrapartum Ultrasound (US) Imaging in Obstetrics and Gynecology (ISUOG) to monitor labour progression through changes in fetal head position. Two reliable ultrasound-derived parameters that are used to predict outcomes of instrumental vaginal delivery are the angle of progression (AoP) and head-symphysis distance (HSD). In this work, as part of the Intrapartum Ultrasounds Grand Challenge (IUGC) 2024, we propose an automated fetal biometry measurement pipeline to reduce intra- and inter-observer variability and improve measurement reliability. Our pipeline consists of three key tasks: (i) classification of standard planes (SP) from US videos, (ii) segmentation of fetal head and pubic symphysis from the detected SPs, and (iii) computation of the AoP and HSD from the segmented regions. We perform sparse sampling to mitigate class imbalances and reduce spurious correlations in task (i), and utilize ensemble-based deep learning methods for task (i) and (ii) to enhance generalizability under different US acquisition settings. Finally, to promote robustness in task iii) with respect to the structural fidelity of measurements, we retain the largest connected components and apply ellipse fitting to the segmentations. Our solution achieved ACC: 0.9452, F1: 0.9225, AUC: 0.983, MCC: 0.8361, DSC: 0.918, HD: 19.73, ASD: 5.71, $\Delta_{AoP}$: 8.90 and $\Delta_{HSD}$: 14.35 across an unseen hold-out set of 4 patients and 224 US frames. The results from the proposed automated pipeline can improve the understanding of labour arrest causes and guide the development of clinical risk stratification tools for efficient and effective prenatal care.

A common neurodegenerative disease, Alzheimer's disease requires a precise diagnosis and efficient treatment, particularly in light of escalating healthcare expenses and the expanding use of artificial intelligence in medical diagnostics. Many recent studies shows that the combination of brain Magnetic Resonance Imaging (MRI) and deep neural networks have achieved promising results for diagnosing AD. Using deep convolutional neural networks, this paper introduces a novel deep learning architecture that incorporates multiresidual blocks, specialized spatial attention blocks, grouped query attention, and multi-head attention. The study assessed the model's performance on four publicly accessible datasets and concentrated on identifying binary and multiclass issues across various categories. This paper also takes into account of the explainability of AD's progression and compared with state-of-the-art methods namely Gradient Class Activation Mapping (GradCAM), Score-CAM, Faster Score-CAM, and XGRADCAM. Our methodology consistently outperforms current approaches, achieving 99.66\% accuracy in 4-class classification, 99.63\% in 3-class classification, and 100\% in binary classification using Kaggle datasets. For Open Access Series of Imaging Studies (OASIS) datasets the accuracies are 99.92\%, 99.90\%, and 99.95\% respectively. The Alzheimer's Disease Neuroimaging Initiative-1 (ADNI-1) dataset was used for experiments in three planes (axial, sagittal, and coronal) and a combination of all planes. The study achieved accuracies of 99.08\% for axis, 99.85\% for sagittal, 99.5\% for coronal, and 99.17\% for all axis, and 97.79\% and 8.60\% respectively for ADNI-2. The network's ability to retrieve important information from MRI images is demonstrated by its excellent accuracy in categorizing AD stages.

Sinonasal pathology can be complex and requires a systematic and meticulous approach. Artificial Intelligence (AI) has the potential to improve diagnostic accuracy and efficiency in sinonasal imaging, but its clinical applicability remains an area of ongoing research. This systematic review evaluates the methodologies and clinical relevance of AI in detecting sinonasal pathology through radiological imaging. Key search terms included "artificial intelligence," "deep learning," "machine learning," "neural network," and "paranasal sinuses,". Abstract and full-text screening was conducted using predefined inclusion and exclusion criteria. Data were extracted on study design, AI architectures used (e.g., Convolutional Neural Networks (CNN), Machine Learning classifiers), and clinical characteristics, such as imaging modality (e.g., Computed Tomography (CT), Magnetic Resonance Imaging (MRI)). A total of 53 studies were analyzed, with 85% retrospective, 68% single-center, and 92.5% using internal databases. CT was the most common imaging modality (60.4%), and chronic rhinosinusitis without nasal polyposis (CRSsNP) was the most studied condition (34.0%). Forty-one studies employed neural networks, with classification as the most frequent AI task (35.8%). Key performance metrics included Area Under the Curve (AUC), accuracy, sensitivity, specificity, precision, and F1-score. Quality assessment based on CONSORT-AI yielded a mean score of 16.0 ± 2. AI shows promise in improving sinonasal imaging interpretation. However, as existing research is predominantly retrospective and single-center, further studies are needed to evaluate AI's generalizability and applicability. More research is also required to explore AI's role in treatment planning and post-treatment prediction for clinical integration.

Microvascular invasion (MVI) is strongly associated with the prognosis of patients with hepatocellular carcinoma (HCC). To evaluate the value of Transformer models with Sonazoid contrast-enhanced ultrasound (CEUS) in the preoperative prediction of MVI. This retrospective study included 164 HCC patients. Deep learning features and radiomic features were extracted from arterial and Kupffer phase images, alongside the collection of clinicopathological parameters. Normality was assessed using the Shapiro-Wilk test. The Mann‒Whitney U-test and least absolute shrinkage and selection operator algorithm were applied to screen features. Transformer, radiomic, and clinical prediction models for MVI were constructed with logistic regression. Repeated random splits followed a 7:3 ratio, with model performance evaluated over 50 iterations. The area under the receiver operating characteristic curve (AUC, 95% confidence interval [CI]), sensitivity, specificity, accuracy, positive predictive value (PPV), negative predictive value (NPV), decision curve, and calibration curve were used to evaluate the performance of the models. The DeLong test was applied to compare performance between models. The Bonferroni method was used to control type I error rates arising from multiple comparisons. A two-sided p-value of < 0.05 was considered statistically significant. In the training set, the diagnostic performance of the arterial-phase Transformer (AT) and Kupffer-phase Transformer (KT) models were better than that of the radiomic and clinical (Clin) models (p < 0.0001). In the validation set, both the AT and KT models outperformed the radiomic and Clin models in terms of diagnostic performance (p < 0.05). The AUC (95% CI) for the AT model was 0.821 (0.72-0.925) with an accuracy of 80.0%, and the KT model was 0.859 (0.766-0.977) with an accuracy of 70.0%. Logistic regression analysis indicated that tumor size (p = 0.016) and alpha-fetoprotein (AFP) (p = 0.046) were independent predictors of MVI. Transformer models using Sonazoid CEUS have potential for effectively identifying MVI-positive patients preoperatively.

Cortical regions such as parietal area H (PH) and the fundus of the superior temporal sulcus (FST) are involved in higher visual function and may play a role in dementia with Lewy bodies (DLB), which is frequently associated with hallucinations. The authors evaluated functional connectivity between these two regions for distinguishing participants with DLB from those with Alzheimer's disease (AD) or mild cognitive impairment (MCI) and from cognitively normal (CN) individuals to identify a functional connectivity MRI signature for DLB. Eighteen DLB participants completed cognitive testing and functional MRI scans and were matched to AD or MCI and CN individuals whose data were obtained from the Alzheimer's Disease Neuroimaging Initiative database (https://adni.loni.usc.edu). Images were analyzed with data from Human Connectome Project (HCP) comparison individuals by using a machine learning-based subject-specific HCP atlas based on diffusion tractography. Bihemispheric functional connectivity of the PH to left FST regions was reduced in the DLB group compared with the AD and CN groups (mean±SD connectivity score=0.307±0.009 vs. 0.456±0.006 and 0.433±0.006, respectively). No significant differences were detected among the groups in connectivity within basal ganglia structures, and no significant correlations were observed between neuropsychological testing results and functional connectivity between the PH and FST regions. Performances on clock-drawing and number-cancelation tests were significantly and negatively correlated with connectivity between the right caudate nucleus and right substantia nigra for DLB participants but not for AD or CN participants. The functional connectivity between PH and FST regions is uniquely affected by DLB and may help distinguish this condition from AD.

Efficient convolutional neural network (CNN) architecture designs have attracted growing research interests. However, they usually apply single receptive field (RF), small asymmetric RFs, or pyramid RFs to learn different feature representations, still encountering two significant challenges in medical image classification tasks: 1) They have limitations in capturing diverse lesion characteristics efficiently, e.g., tiny, coordination, small and salient, which have unique roles on results, especially imbalanced medical image classification. 2) The predictions generated by those CNNs are often unfair/biased, bringing a high risk by employing them to real-world medical diagnosis conditions. To tackle these issues, we develop a new concept, Expert-Like Reparameterization of Heterogeneous Pyramid Receptive Fields (ERoHPRF), to simultaneously boost medical image classification performance and fairness. This concept aims to mimic the multi-expert consultation mode by applying the well-designed heterogeneous pyramid RF bags to capture different lesion characteristics effectively via convolution operations with multiple heterogeneous kernel sizes. Additionally, ERoHPRF introduces an expert-like structural reparameterization technique to merge its parameters with the two-stage strategy, ensuring competitive computation cost and inference speed through comparisons to a single RF. To manifest the effectiveness and generalization ability of ERoHPRF, we incorporate it into mainstream efficient CNN architectures. The extensive experiments show that our method maintains a better trade-off than state-of-the-art methods in terms of medical image classification, fairness, and computation overhead. The codes of this paper will be released soon.

To establish and validate deep learning (DL) models based on pre-treatment multiparametric magnetic resonance imaging (MRI) images of primary rectal cancer and basic clinical data for the prediction of synchronous liver metastases (SLM) in patients with Rectal cancer (RC). In this retrospective study, 176 and 31 patients with RC who underwent multiparametric MRI from two centers were enrolled in the primary and external validation cohorts, respectively. Clinical factors, including sex, primary tumor site, CEA level, and CA199 level were assessed. A clinical feature (CF) model was first developed by multivariate logistic regression, then two residual network DL models were constructed based on multiparametric MRI of primary cancer with or without CF incorporation. Finally, the SLM prediction models were validated by 5-fold cross-validation and external validation. The performance of the models was evaluated by decision curve analysis (DCA) and receiver operating characteristic (ROC) analysis. Among three SLM prediction models, the Combined DL model integrating primary tumor MRI and basic clinical data achieved the best performance (AUC = 0.887 in primary study cohort; AUC = 0.876 in the external validation cohort). In the primary study cohort, the CF model, MRI DL model, and Combined DL model achieved AUCs of 0.816 (95% CI: 0.750, 0.881), 0.788 (95% CI: 0.720, 0.857), and 0.887 (95% CI: 0.834, 0.940) respectively. In the external validation cohort, the CF model, DL model without CF, and DL model with CF achieved AUCs of 0.824 (95% CI: 0.664, 0.984), 0.662 (95% CI: 0.461, 0.863), and 0.876 (95% CI: 0.728, 1.000), respectively. The combined DL model demonstrates promising potential to predict SLM in patients with RC, thereby making individualized imaging test strategies. Accurate synchronous liver metastasis (SLM) risk stratification is important for treatment planning and prognosis improvement. The proposed DL signature may be employed to better understand an individual patient's SLM risk, aiding in treatment planning and selection of further imaging examinations to personalize clinical decisions. Not applicable.

To develop and validate a deep learning system with guided diffusion-based data augmentation for grading partial-thickness supraspinatus tendon (SST) tears and to compare its performance with experienced radiologists, including external validation. This retrospective study included 1150 patients with arthroscopically confirmed SST tears, divided into a training set (741 patients), validation set (185 patients), and internal test set (185 patients). An independent external test set of 224 patients was used for generalizability assessment. To address data imbalance, MRI images were augmented using a guided diffusion model. A ResNet-34 model was employed for Ellman grading of bursal-sided and articular-sided partial-thickness tears across different MRI sequences (oblique coronal [OCOR], oblique sagittal [OSAG], and combined OCOR+OSAG). Performance was evaluated using AUC and precision-recall curves, and compared to three experienced musculoskeletal (MSK) radiologists. The DeLong test was used to compare performance across different sequence combinations. A total of 26,020 OCOR images and 26,356 OSAG images were generated using the guided diffusion model. For bursal-sided partial-thickness tears in the internal dataset, the model achieved AUCs of 0.99, 0.98, and 0.97 for OCOR, OSAG, and combined sequences, respectively, while for articular-sided tears, AUCs were 0.99, 0.99, and 0.99. The DeLong test showed no significant differences among sequence combinations (P=0.17, 0.14, 0.07). In the external dataset, the combined-sequence model achieved AUCs of 0.99, 0.97, and 0.97 for bursal-sided tears and 0.99, 0.95, and 0.95 for articular-sided tears. Radiologists demonstrated an ICC of 0.99, but their grading performance was significantly lower than the ResNet-34 model (P<0.001). The deep learning system improved grading consistency and significantly reduced evaluation time, while guided diffusion augmentation enhanced model robustness. The proposed deep learning system provides a reliable and efficient method for grading partial-thickness SST tears, achieving radiologist-level accuracy with greater consistency and faster evaluation speed.