Existing saliency-guided training approaches improve model generalization by incorporating a loss term that compares the model's class activation map (CAM) for a sample's true-class ({\it i.e.}, correct-label class) against a human reference saliency map. However, prior work has ignored the false-class CAM(s), that is the model's saliency obtained for incorrect-label class. We hypothesize that in binary tasks the true and false CAMs should diverge on the important classification features identified by humans (and reflected in human saliency maps). We use this hypothesis to motivate three new saliency-guided training methods incorporating both true- and false-class model's CAM into the training strategy and a novel post-hoc tool for identifying important features. We evaluate all introduced methods on several diverse binary close-set and open-set classification tasks, including synthetic face detection, biometric presentation attack detection, and classification of anomalies in chest X-ray scans, and find that the proposed methods improve generalization capabilities of deep learning models over traditional (true-class CAM only) saliency-guided training approaches. We offer source codes and model weights\footnote{GitHub repository link removed to preserve anonymity} to support reproducible research.

Alzheimer's disease (AD) is a progressive neurodegenerative disorder that predominantly affects the elderly population and currently has no cure. Magnetic Resonance Imaging (MRI), as a non-invasive imaging technique, is essential for the early diagnosis of AD. MRI inherently contains both spatial and frequency information, as raw signals are acquired in the frequency domain and reconstructed into spatial images via the Fourier transform. However, most existing AD diagnostic models extract features from a single domain, limiting their capacity to fully capture the complex neuroimaging characteristics of the disease. While some studies have combined spatial and frequency information, they are mostly confined to 2D MRI, leaving the potential of dual-domain analysis in 3D MRI unexplored. To overcome this limitation, we propose Spatio-Frequency Network (SFNet), the first end-to-end deep learning framework that simultaneously leverages spatial and frequency domain information to enhance 3D MRI-based AD diagnosis. SFNet integrates an enhanced dense convolutional network to extract local spatial features and a global frequency module to capture global frequency-domain representations. Additionally, a novel multi-scale attention module is proposed to further refine spatial feature extraction. Experiments on the Alzheimer's Disease Neuroimaging Initiative (ADNI) dataset demonstrate that SFNet outperforms existing baselines and reduces computational overhead in classifying cognitively normal (CN) and AD, achieving an accuracy of 95.1%.

Cardiac ultrasound exams provide real-time data to guide clinical decisions but require highly trained sonographers. Artificial intelligence (AI) that uses deep learning algorithms to guide novices in the acquisition of diagnostic echocardiographic studies may broaden access and improve care. The objective of this trial was to evaluate whether nurses without previous ultrasound experience (novices) could obtain diagnostic-quality acquisitions of 10 echocardiographic views using AI-based software. This noninferiority study was prospective, international, nonrandomized, and conducted at 2 medical centers, in the United States and France, from November 2023 to August 2024. Two limited cardiac exams were performed on adult patients scheduled for a clinically indicated echocardiogram; one was conducted by a novice using AI guidance and one by an expert (experienced sonographer or cardiologist) without it. Primary endpoints were evaluated by 5 experienced cardiologists to assess whether the novice exam was of sufficient quality to visually analyze the left ventricular size and function, the right ventricle size, and the presence of nontrivial pericardial effusion. Secondary endpoints included 8 additional cardiac parameters. A total of 240 patients (mean age 62.6 years; 117 women (48.8%); mean body mass index 26.6 kg/m²) completed the study. One hundred percent of the exams performed by novices with the studied software were of sufficient quality to assess the primary endpoints. Cardiac parameters assessed in exams conducted by novices and experts were strongly correlated. AI-based software provides a safe means for novices to perform diagnostic-quality cardiac ultrasounds after a short training period.

Precise segmentation of papillary thyroid microcarcinoma (PTMC) during ultrasound-guided radiofrequency ablation (RFA) is critical for effective treatment but remains challenging due to acoustic artifacts, small lesion size, and anatomical variability. In this study, we propose DualSwinUnet++, a dual-decoder transformer-based architecture designed to enhance PTMC segmentation by incorporating thyroid gland context. DualSwinUnet++ employs independent linear projection heads for each decoder and a residual information flow mechanism that passes intermediate features from the first (thyroid) decoder to the second (PTMC) decoder via concatenation and transformation. These design choices allow the model to condition tumor prediction explicitly on gland morphology without shared gradient interference. Trained on a clinical ultrasound dataset with 691 annotated RFA images and evaluated against state-of-the-art models, DualSwinUnet++ achieves superior Dice and Jaccard scores while maintaining sub-200ms inference latency. The results demonstrate the model's suitability for near real-time surgical assistance and its effectiveness in improving segmentation accuracy in challenging PTMC cases.

Myocardial strain plays a crucial role in diagnosing heart failure and myocardial infarction. Its computation relies on assessing heart muscle motion throughout the cardiac cycle. This assessment can be performed by following key points on each frame of a cine Magnetic Resonance Imaging (MRI) sequence. The use of segmentation labels yields more accurate motion estimation near heart muscle boundaries. However, since few frames in a cardiac sequence usually have segmentation labels, most methods either rely on annotated pairs of frames/volumes, greatly reducing available data, or use all frames of the cardiac cycle without segmentation supervision. Moreover, these techniques rarely utilize more than two phases during training. In this work, a new semi-supervised motion estimation algorithm using all frames of the cardiac sequence is presented. The distance map generated from the end-diastolic segmentation label is used to weight loss functions. The method is tested on an in-house dataset containing 271 patients. Several deep learning image registration and tracking algorithms were retrained on our dataset and compared to our approach. The proposed approach achieves an average End Point Error (EPE) of 1.02mm, against 1.19mm for RAFT (Recurrent All-Pairs Field Transforms). Using the end-diastolic distance map further improves this metric to 0.95mm compared to 0.91 for the fully supervised version. Correlations in systolic peak were 0.83 and 0.90 for the left ventricular global radial and circumferential strain respectively, and 0.91 for the right ventricular circumferential strain.

Accurate and efficient medical image segmentation is crucial but challenging due to anatomical variability and high computational demands on volumetric data. Recent hybrid CNN-Transformer architectures achieve state-of-the-art results but add significant complexity. In this paper, we propose MLRU++, a Multiscale Lightweight Residual UNETR++ architecture designed to balance segmentation accuracy and computational efficiency. It introduces two key innovations: a Lightweight Channel and Bottleneck Attention Module (LCBAM) that enhances contextual feature encoding with minimal overhead, and a Multiscale Bottleneck Block (M2B) in the decoder that captures fine-grained details via multi-resolution feature aggregation. Experiments on four publicly available benchmark datasets (Synapse, BTCV, ACDC, and Decathlon Lung) demonstrate that MLRU++ achieves state-of-the-art performance, with average Dice scores of 87.57% (Synapse), 93.00% (ACDC), and 81.12% (Lung). Compared to existing leading models, MLRU++ improves Dice scores by 5.38% and 2.12% on Synapse and ACDC, respectively, while significantly reducing parameter count and computational cost. Ablation studies evaluating LCBAM and M2B further confirm the effectiveness of the proposed architectural components. Results suggest that MLRU++ offers a practical and high-performing solution for 3D medical image segmentation tasks. Source code is available at: https://github.com/1027865/MLRUPP

Faithfulness and interpretability are essential for deploying deep neural networks (DNNs) in safety-critical domains such as medical imaging. B-cos networks offer a promising solution by replacing standard linear layers with a weight-input alignment mechanism, producing inherently interpretable, class-specific explanations without post-hoc methods. While maintaining diagnostic performance competitive with state-of-the-art DNNs, standard B-cos models suffer from severe aliasing artifacts in their explanation maps, making them unsuitable for clinical use where clarity is essential. In this work, we address these limitations by introducing anti-aliasing strategies using FLCPooling (FLC) and BlurPool (BP) to significantly improve explanation quality. Our experiments on chest X-ray datasets demonstrate that the modified $\text{B-cos}_\text{FLC}$ and $\text{B-cos}_\text{BP}$ preserve strong predictive performance while providing faithful and artifact-free explanations suitable for clinical application in multi-class and multi-label settings. Code available at: GitHub repository (url: https://github.com/mkleinma/B-cos-medical-paper).

Magnetic resonance imaging (MRI) involves a trade-off between imaging time, signal-to-noise ratio (SNR), and spatial resolution. Reducing the imaging time often leads to a lower SNR or resolution. Deep-learning-based reconstruction (DLR) methods have been introduced to address these limitations. Image-domain super-resolution DLR enables high resolution without additional image scans. High-quality images can be obtained within a shorter timeframe by appropriately configuring DLR parameters. It is necessary to maximize the performance of super-resolution DLR to enable efficient use in MRI. We evaluated the performance of a vendor-provided super-resolution DLR method (PIQE) on a Canon 3 T MRI scanner using an edge phantom and clinical brain images from eight patients. Quantitative assessment included structural similarity index (SSIM), peak SNR (PSNR), root mean square error (RMSE), and full width at half maximum (FWHM). FWHM was used to quantitatively assess spatial resolution and image sharpness. Visual evaluation using a five-point Likert scale was also performed to assess perceived image quality. Image domain super-resolution DLR reduced scan time by up to 70 % while preserving the structural image quality. Acquisition matrices of 0.87 mm/pixel or finer with a zoom ratio of ×2 yielded SSIM ≥0.80, PSNR ≥35 dB, and non-significant FWHM differences compared to full-resolution references. In contrast, aggressive downsampling (zoom ratio 3 from low-resolution matrices) led to image degradation including truncation artifacts and reduced sharpness. These results clarify the optimal use of PIQE as an image-domain super-resolution method and provide practical guidance for its application in clinical MRI workflows.

Parkinson's disease (PD) and progressive supranuclear palsy (PSP) present similar clinical symptoms, but their treatment options and clinical prognosis differ significantly. Therefore, we aimed to discriminate between PD and PSP based on multi-level indices of resting-state functional magnetic resonance imaging (rs-fMRI) via the machine learning approach. A total of 58 PD and 52 PSP patients were prospectively enrolled in this study. Participants were randomly allocated to a training set and a validation set in a 7:3 ratio. Various rs-fMRI indices were extracted, followed by a comprehensive feature screening for each index. We constructed fifteen distinct combinations of indices and selected four machine learning algorithms for model development. Subsequently, different validation templates were employed to assess the classification results and investigate the relationship between the most significant features and clinical assessment scales. The classification performance of logistic regression (LR) and support vector machine (SVM) models, based on multiple index combinations, was significantly superior to that of other machine learning models and combinations when utilizing automatic anatomical labeling (AAL) templates. This has been verified across different templates. The utilization of multiple rs-fMRI indices significantly enhances the performance of machine learning models and can effectively achieve the automatic identification of PD and PSP at the individual level.

Thyroid eye disease (TED) is a common, complex orbital disorder characterized by soft-tissue changes visible on imaging. Artificial intelligence (AI) offers promises for improving TED diagnosis and treatment; however, no systematic review has yet characterized the research landscape, key challenges, and future directions. We followed PRISMA guidelines to search multiple databases until January, 2025, for studies applying AI to computed tomography (CT), magnetic resonance imaging, and nuclear, facial or retinal imaging in TED patients. Using the APPRAISE-AI tool, we assessed study quality and included 41 studies covering various AI applications. Sample sizes ranged from 33 to 2,288 participants, predominantly East Asian. CT and facial imaging were the most common modalities, reported in 16 and 13 articles, respectively. Studies addressed clinical tasks-diagnosis, activity assessment, severity grading, and treatment prediction-and technical tasks-classification, segmentation, and image generation-with classification being the most frequent. Researchers primarily employed deep-learning models, such as residual network (ResNet) and Visual Geometry Group (VGG). Overall, the majority of the studies were of moderate quality. Image-based AI shows strong potential to improve diagnostic accuracy and guide personalized treatment strategies in TED. Future research should prioritize robust study designs, the creation of public datasets, multimodal imaging integration, and interdisciplinary collaboration to accelerate clinical translation.