To develop and evaluate a hierarchical deep learning system that detects orbital fractures on computed tomography (CT) images and classifies them as depressed or trap-door types. A retrospective diagnostic accuracy study analyzing CT images from patients with confirmed orbital fractures. We collected CT images from 686 patients with orbital fractures treated at a single institution (2010-2025), resulting in 46,013 orbital CT slices. After preprocessing, 7809 slices were selected as regions of interest and partitioned into training (6508 slices) and test (1301 slices) datasets. Our hierarchical approach consisted of a first-stage classifier (YOLOv8) for fracture detection and a second-stage classifier (Vision Transformer) for distinguishing depressed from trap-door fractures. Performance was evaluated at both slice and patient levels, focusing on accuracy, sensitivity, specificity, and area under the receiver operating characteristic curve (AUC-ROC) at both slice and patient levels. For fracture detection, YOLOv8 achieved a slice-level sensitivity of 80.4 % and specificity of 79.2 %, with patient-level performance improving to 94.7 % sensitivity and 90.0 % specificity. For fracture classification, Vision Transformer demonstrated a slice-level sensitivity of 91.5 % and specificity of 83.5 % for trap-door and depressed fractures, with patient-level metrics of 100 % sensitivity and 88.9 % specificity. The complete system correctly identified 18/20 no-fracture cases, 35/40 depressed fracture cases, and 15/17 trap-door fracture cases. Our hierarchical deep learning system effectively detects orbital fractures and distinguishes between depressed and trap-door types with high accuracy. This approach could aid in the timely identification of trap-door fractures requiring urgent surgical intervention, particularly in settings lacking specialized expertise.

While total intracranial carotid artery calcification (ICAC) volume is an established stroke biomarker, growing evidence shows this aggregate metric ignores the critical influence of plaque location, since calcification in different segments carries distinct prognostic and procedural risks. However, a finer-grained, segment-specific quantification has remained technically infeasible. Conventional 3D models are forced to process downsampled volumes or isolated patches, sacrificing the global context required to resolve anatomical ambiguity and render reliable landmark localization. To overcome this, we reformulate the 3D challenge as a \textbf{Parallel Probabilistic Landmark Localization} task along the 1D axial dimension. We propose the \textbf{Depth-Sequence Transformer (DST)}, a framework that processes full-resolution CT volumes as sequences of 2D slices, learning to predict $N=6$ independent probability distributions that pinpoint key anatomical landmarks. Our DST framework demonstrates exceptional accuracy and robustness. Evaluated on a 100-patient clinical cohort with rigorous 5-fold cross-validation, it achieves a Mean Absolute Error (MAE) of \textbf{0.1 slices}, with \textbf{96\%} of predictions falling within a $\pm1$ slice tolerance. Furthermore, to validate its architectural power, the DST backbone establishes the best result on the public Clean-CC-CCII classification benchmark under an end-to-end evaluation protocol. Our work delivers the first practical tool for automated segment-specific ICAC analysis. The proposed framework provides a foundation for further studies on the role of location-specific biomarkers in diagnosis, prognosis, and procedural planning. Our code will be made publicly available.

In medical image segmentation, convolutional neural networks (CNNs) and transformers are dominant. For CNNs, given the local receptive fields of convolutional layers, long-range spatial correlations are captured through consecutive convolutions and pooling. However, as the computational cost and memory footprint can be prohibitively large, 3D models can only afford fewer layers than 2D models with reduced receptive fields and abstract levels. For transformers, although long-range correlations can be captured by multi-head attention, its quadratic complexity with respect to input size is computationally demanding. Therefore, either model may require input size reduction to allow more filters and layers for better segmentation. Nevertheless, given their discrete nature, models trained with patch-wise training or image downsampling may produce suboptimal results when applied on higher resolutions. To address this issue, here we propose the resolution-robust HNOSeg-XS architecture. We model image segmentation by learnable partial differential equations through the Fourier neural operator which has the zero-shot super-resolution property. By replacing the Fourier transform by the Hartley transform and reformulating the problem in the frequency domain, we created the HNOSeg-XS model, which is resolution robust, fast, memory efficient, and extremely parameter efficient. When tested on the BraTS'23, KiTS'23, and MVSeg'23 datasets with a Tesla V100 GPU, HNOSeg-XS showed its superior resolution robustness with fewer than 34.7k model parameters. It also achieved the overall best inference time (< 0.24 s) and memory efficiency (< 1.8 GiB) compared to the tested CNN and transformer models.

The labor-intensive nature of medical data annotation presents a significant challenge for respiratory disease diagnosis, resulting in a scarcity of high-quality labeled datasets in resource-constrained settings. Moreover, patient privacy concerns complicate the direct sharing of local medical data across institutions, and existing centralized data-driven approaches, which rely on amounts of available data, often compromise data privacy. This study proposes a federated few-shot learning framework with privacy-preserving mechanisms to address the issues of limited labeled data and privacy protection in diagnosing respiratory diseases. In particular, a meta-stochastic gradient descent algorithm is proposed to mitigate the overfitting problem that arises from insufficient data when employing traditional gradient descent methods for neural network training. Furthermore, to ensure data privacy against gradient leakage, differential privacy noise from a standard Gaussian distribution is integrated into the gradients during the training of private models with local data, thereby preventing the reconstruction of medical images. Given the impracticality of centralizing respiratory disease data dispersed across various medical institutions, a weighted average algorithm is employed to aggregate local diagnostic models from different clients, enhancing the adaptability of a model across diverse scenarios. Experimental results show that the proposed method yields compelling results with the implementation of differential privacy, while effectively diagnosing respiratory diseases using data from different structures, categories, and distributions.

Recent works have revisited the infamous task ``Name That Dataset'', demonstrating that non-medical datasets contain underlying biases and that the dataset origin task can be solved with high accuracy. In this work, we revisit the same task applied to popular open-source chest X-ray datasets. Medical images are naturally more difficult to release for open-source due to their sensitive nature, which has led to certain open-source datasets being extremely popular for research purposes. By performing the same task, we wish to explore whether dataset bias also exists in these datasets. To extend our work, we apply simple transformations to the datasets, repeat the same task, and perform an analysis to identify and explain any detected biases. Given the importance of AI applications in medical imaging, it's vital to establish whether modern methods are taking shortcuts or are focused on the relevant pathology. We implement a range of different network architectures on the datasets: NIH, CheXpert, MIMIC-CXR and PadChest. We hope this work will encourage more explainable research being performed in medical imaging and the creation of more open-source datasets in the medical domain. Our code can be found here: https://github.com/eedack01/x_ray_ds_bias.

Lacunes, which are small fluid-filled cavities in the brain, are signs of cerebral small vessel disease and have been clinically associated with various neurodegenerative and cerebrovascular diseases. Hence, accurate detection of lacunes is crucial and is one of the initial steps for the precise diagnosis of these diseases. However, developing a robust and consistently reliable method for detecting lacunes is challenging because of the heterogeneity in their appearance, contrast, shape, and size. To address the above challenges, in this study, we propose a lacune detection method using the Segment Anything Model (SAM), guided by point prompts from a candidate prompt generator. The prompt generator initially detects potential lacunes with a high sensitivity using a composite loss function. The SAM model selects true lacunes by delineating their characteristics from mimics such as the sulcus and enlarged perivascular spaces, imitating the clinicians strategy of examining the potential lacunes along all three axes. False positives were further reduced by adaptive thresholds based on the region-wise prevalence of lacunes. We evaluated our method on two diverse, multi-centric MRI datasets, VALDO and ISLES, comprising only FLAIR sequences. Despite diverse imaging conditions and significant variations in slice thickness (0.5-6 mm), our method achieved sensitivities of 84% and 92%, with average false positive rates of 0.05 and 0.06 per slice in ISLES and VALDO datasets respectively. The proposed method outperformed the state-of-the-art methods, demonstrating its effectiveness in lacune detection and quantification.

Recent works have revisited the infamous task ``Name That Dataset'', demonstrating that non-medical datasets contain underlying biases and that the dataset origin task can be solved with high accuracy. In this work, we revisit the same task applied to popular open-source chest X-ray datasets. Medical images are naturally more difficult to release for open-source due to their sensitive nature, which has led to certain open-source datasets being extremely popular for research purposes. By performing the same task, we wish to explore whether dataset bias also exists in these datasets. To extend our work, we apply simple transformations to the datasets, repeat the same task, and perform an analysis to identify and explain any detected biases. Given the importance of AI applications in medical imaging, it's vital to establish whether modern methods are taking shortcuts or are focused on the relevant pathology. We implement a range of different network architectures on the datasets: NIH, CheXpert, MIMIC-CXR and PadChest. We hope this work will encourage more explainable research being performed in medical imaging and the creation of more open-source datasets in the medical domain. Our code can be found here: https://github.com/eedack01/x_ray_ds_bias.

Brucella spondylitis (BS) and tuberculous spondylitis (TS) are prevalent spinal infections with distinct treatment protocols. Rapid and accurate differentiation between these two conditions is crucial for effective clinical management; however, current imaging and pathogen-based diagnostic methods fall short of fully meeting clinical requirements. This study explores the feasibility of employing deep learning (DL) models based on conventional magnetic resonance imaging (MRI) to differentiate BS and TS. A total of 310 subjects were enrolled in our hospital, comprising 209 with BS, 101 with TS. The participants were randomly divided into a training set (n = 217) and a test set (n = 93). And 74 with other hospital was external validation set. Integrating Convolutional Block Attention Module (CBAM) into the ResNeXt-50 architecture and training the model using sagittal T2-weighted images (T2WI). Classification performance was evaluated using the area under the receiver operating characteristic (AUC) curve, and diagnostic accuracy was compared against general models such as ResNet50, GoogleNet, EfficientNetV2, and VGG16. The CBAM-ResNeXt model revealed superior performance, with accuracy, precision, recall, F1-score, and AUC from 0.942, 0.940, 0.928, 0.934, 0.953, respectively. These metrics outperformed those of the general models. The proposed model offers promising potential for the diagnosis of BS and TS using conventional MRI. It could serve as an invaluable tool in clinical practice, providing a reliable reference for distinguishing between these two diseases.

In most medical centers, particularly in primary hospitals, non-contrast computed tomography (NCCT) serves as the primary imaging modality for diagnosing acute ischemic stroke. However, due to the small density difference between the infarct and the surrounding normal brain tissue on NCCT images within the initial 6 h post-onset, it poses significant challenges in promptly and accurately positioning and quantifying the infarct at the early stage. To investigate whether a radiomics-based model using NCCT could effectively assess the risk of acute ischemic stroke (AIS). This study proposed a machine learning (ML) for infarct detection, enabling automated quantitative assessment of AIS lesions on NCCT images. In this retrospective study, NCCT images from 228 patients with AIS (< 6 h from onset) were included, and paired with MRI-diffusion-weighted imaging (DWI) images (attained within 1 to 7 days of onset). NCCT and DWI images were co-registered using the Elastix toolbox. The internal dataset (153 AIS patients) included 179 AIS VOIs and 153 non-AIS VOIs as the training and validation groups. Subsequent cases (75 patients) after 2021 served as the independent test set, comprising 94 AIS VOIs and 75 non-AIS VOIs. The random forest (RF) model demonstrated robust diagnostic performance across the training, validation, and independent test sets. The areas under the receiver operating characteristic (ROC) curves were 0.858 (95% CI: 0.808-0.908), 0.829 (95% CI: 0.748-0.910), and 0.789 (95% CI: 0.717-0.860), respectively. Accuracies were 79.399%, 77.778%, and 73.965%, while sensitivities were 81.679%, 77.083%, and 68.085%. Specificities were 76.471%, 78.431%, and 81.333%, respectively. NCCT-based radiomics combined with a machine learning model could discriminate between AIS and non-AIS patients within less than 6 h of onset. This approach holds promise for improving early stroke diagnosis and patient outcomes. Not applicable.

Tonsilloliths can be seen on panoramic radiographs (PRs) as deposits located on the middle portion of the ramus of the mandible. Although tonsilloliths are clinically harmless, the high risk of misdiagnosis leads to unnecessary advanced examinations and interventions, thus jeopardizing patient safety and increasing unnecessary resource use in the healthcare system. Therefore, this study aims to meet an important clinical need by providing accurate and rapid diagnostic support. The dataset consisted of a total of 275 PRs, with 125 PRs lacking tonsillolith and 150 PRs having tonsillolith. ResNet and EfficientNet CNN models were assessed during the model selection process. An evaluation was conducted to analyze the learning capacity, intricacy, and compatibility of each model with the problem at hand. The effectiveness of the models was evaluated using accuracy, recall, precision, and F1 score measures following the training phase. Both the ResNet18 and EfficientNetB0 models were able to differentiate between tonsillolith-present and tonsillolith-absent conditions with an average accuracy of 89%. ResNet101 demonstrated underperformance when contrasted with other models. EfficientNetB1 exhibits satisfactory accuracy in both categories. The EfficientNetB0 model exhibits a 93% precision, 87% recall, 90% F1 score, and 89% accuracy. This study indicates that implementing AI-powered deep learning techniques would significantly improve the clinical diagnosis of tonsilloliths.