Detecting malignant transformation of sinonasal inverted papilloma (SIP) into squamous cell carcinoma (SIP-SCC) before surgery is a clinical need. We aimed to explore the value of deep learning (DL) that leverages nasal endoscopy and T2-weighted magnetic resonance imaging (T2W-MRI) for automated tumor segmentation and differentiation between SIP and SIP-SCC. We conducted a retrospective analysis of 174 patients diagnosed with SIPs, who were divided into a training cohort (n = 121) and a testing cohort (n = 53). Three DL architectures were utilized to train automated segmentation models for endoscopic and T2W-MRI images. DL scores predicting SIP-SCC were generated using DenseNet121 from both modalities and combined to create a dual-modality DL nomogram. The diagnostic performance of the DL models was assessed alongside two radiologists, evaluated through the area under the receiver operating characteristic curve (AUROC), with comparisons made using the Delong method. In the testing cohort, the FCN_ResNet101 and VNet exhibited superior performance in automated segmentation, achieving mean dice similarity coefficients of 0.95 ± 0.03 for endoscopy and 0.93 ± 0.02 for T2W-MRI, respectively. The dual-modality DL nomogram based on automated segmentation demonstrated the highest predictive performance for SIP-SCC (AUROC 0.865), outperforming the radiology resident (AUROC 0.672, p = 0.071) and the attending radiologist (AUROC 0.707, p = 0.066), with a trend toward significance. Notably, both radiologists improved their diagnostic performance with the assistance of the DL nomogram (AUROCs 0.734 and 0.834). The DL framework integrating endoscopy and T2W-MRI offers a fully automated predictive tool for SIP-SCC. The integration of endoscopy and T2W-MRI within a well-established DL framework enables fully automated prediction of SIP-SSC, potentially improving decision-making for patients with suspicious SIP. Detecting the transformation of SIP into SIP-SCC before surgery is both critical and challenging. Endoscopy and T2W-MRI were integrated using DL for predicting SIP-SCC. The dual-modality DL nomogram outperformed two radiologists. The nomogram may improve decision-making for patients with suspicious SIP.

To validate the diagnostic performance of a B-mode ultrasound-based deep learning (DL) model in distinguishing benign and malignant cervical lymphadenopathy (CLP). A total of 210 CLPs with conclusive pathological results were retrospectively included and separated as training (n = 169) or test cohort (n = 41) randomly at a ratio of 4:1. A DL model integrating convolutional neural network, deformable convolution network and attention mechanism was developed. Three diagnostic models were developed: (a) Model I, CLPs with at least one suspicious B-mode ultrasound feature (ratio of longitudinal to short diameter < 2, irregular margin, hyper-echogenicity, hilus absence, cystic necrosis and calcification) were deemed malignant; (b) Model II: total risk score of B-mode ultrasound features obtained by multivariate logistic regression and (c) Model III: CLPs with positive DL output are deemed malignant. The diagnostic utility of these models was assessed by the area under the receiver operating curve (AUC) and corresponding sensitivity and specificity. Multivariate analysis indicated that DL positive result was the most important factor associated with malignant CLPs [odds ratio (OR) = 39.05, p < 0.001], only followed by hilus absence (OR = 6.01, p = 0.001) in the training cohort. In the test cohort, the AUC of the DL model (0.871) was significantly higher than that in model I (AUC = 0.681, p = 0.04) and model II (AUC = 0.679, p = 0.03), respectively. In addition, model III obtained 93.3% specificity, which was significantly higher than that in model I (40.0%, p = 0.002) and model II (60.0%, p = 0.03), respectively. Although the sensitivity of model I was the highest, it did not show a significant difference compared to that of model III (96.2% vs.80.8%, p = 0.083). B-mode ultrasound-based DL is a potentially robust tool for the differential diagnosis of benign and malignant CLPs.

In convolutional neural networks (CNNs), downsampling operations are crucial to model performance. Although traditional downsampling methods (such as maximum pooling and cross-row convolution) perform well in feature aggregation, receptive field expansion, and computational reduction, they may lead to the loss of key spatial information in semantic segmentation tasks, thereby affecting the pixel-by-pixel prediction accuracy.To this end, this study proposes a downsampling method based on information complementarity - Hybrid Pooling Downsampling (HPD). The core is to replace the traditional method with MinMaxPooling, and effectively retain the light and dark contrast and detail features of the image by extracting the maximum value information of the local area.Experiment on various CNN architectures on the ACDC and Synapse datasets show that HPD outperforms traditional methods in segmentation performance, and increases the DSC coefficient by 0.5% on average. The results show that the HPD module provides an efficient solution for semantic segmentation tasks.

Medical radiology reports play a crucial role in diagnosing various diseases, yet generating them manually is time-consuming and burdens clinical workflows. Medical radiology report generation aims to automate this process using deep learning to assist radiologists and reduce patient wait times. This study presents the most comprehensive systematic review to date on deep learning-based MRRG, encompassing recent advances that span traditional architectures to large language models. We focus on available datasets, modeling approaches, and evaluation practices. Following PRISMA guidelines, we retrieved 323 articles from major academic databases and included 78 studies after eligibility screening. We critically analyze key components such as model architectures, loss functions, datasets, evaluation metrics, and optimizers - identifying 22 widely used datasets, 14 evaluation metrics, around 20 loss functions, over 25 visual backbones, and more than 30 textual backbones. To support reproducibility and accelerate future research, we also compile links to modern models, toolkits, and pretrained resources. Our findings provide technical insights and outline future directions to address current limitations, promoting collaboration at the intersection of medical imaging, natural language processing, and deep learning to advance trustworthy AI systems in radiology.

Recent advances in medical image segmentation have been driven by deep learning; however, most existing methods remain limited by modality-specific designs and exhibit poor adaptability to dynamic medical imaging scenarios. The Segment Anything Model 2 (SAM2) and its related variants, which introduce a streaming memory mechanism for real-time video segmentation, present new opportunities for prompt-based, generalizable solutions. Nevertheless, adapting these models to medical video scenarios typically requires large-scale datasets for retraining or transfer learning, leading to high computational costs and the risk of catastrophic forgetting. To address these challenges, we propose DD-SAM2, an efficient adaptation framework for SAM2 that incorporates a Depthwise-Dilated Adapter (DD-Adapter) to enhance multi-scale feature extraction with minimal parameter overhead. This design enables effective fine-tuning of SAM2 on medical videos with limited training data. Unlike existing adapter-based methods focused solely on static images, DD-SAM2 fully exploits SAM2's streaming memory for medical video object tracking and segmentation. Comprehensive evaluations on TrackRad2025 (tumor segmentation) and EchoNet-Dynamic (left ventricle tracking) datasets demonstrate superior performance, achieving Dice scores of 0.93 and 0.97, respectively. To the best of our knowledge, this work provides an initial attempt at systematically exploring adapter-based SAM2 fine-tuning for medical video segmentation and tracking. Code, datasets, and models will be publicly available at https://github.com/apple1986/DD-SAM2.

Lymphoma poses a critical health challenge worldwide, demanding computer aided solutions towards diagnosis, treatment, and research to significantly enhance patient outcomes and combat this pervasive disease. Accurate classification of lymphoma subtypes from Whole Slide Images (WSIs) remains a complex challenge due to morphological similarities among subtypes and the limitations of models that fail to jointly capture local and global features. Traditional diagnostic methods, limited by subjectivity and inconsistencies, highlight the need for advanced, Artificial Intelligence (AI)-driven solutions. This study proposes a hybrid deep learning framework-Hybrid Convolutional and Transformer Network for Lymphoma Classification (HCTN-LC)-designed to enhance the precision and interpretability of lymphoma subtype classification. The model employs a dual-pathway architecture that combines a lightweight SqueezeNet for local feature extraction with a Vision Transformer (ViT) for capturing global context. A Feature Fusion and Enhancement Module (FFEM) is introduced to dynamically integrate features from both pathways. The model is trained and evaluated on a large WSI dataset encompassing three lymphoma subtypes: CLL, FL, and MCL. HCTN-LC achieves superior performance with an overall accuracy of 99.87%, sensitivity of 99.87%, specificity of 99.93%, and AUC of 0.9991, outperforming several recent hybrid models. Grad-CAM visualizations confirm the model's focus on diagnostically relevant regions. The proposed HCTN-LC demonstrates strong potential for real-time and low-resource clinical deployment, offering a robust and interpretable AI tool for hematopathological diagnosis.

Deep learning models often hallucinate, producing realistic artifacts that are not truly present in the sample. This can have dire consequences for scientific and medical inverse problems, such as MRI and microscopy denoising, where accuracy is more important than perceptual quality. Uncertainty quantification techniques, such as conformal prediction, can pinpoint outliers and provide guarantees for image regression tasks, improving reliability. However, existing methods utilize a linear constant scaling factor to calibrate uncertainty bounds, resulting in larger, less informative bounds. We propose QUTCC, a quantile uncertainty training and calibration technique that enables nonlinear, non-uniform scaling of quantile predictions to enable tighter uncertainty estimates. Using a U-Net architecture with a quantile embedding, QUTCC enables the prediction of the full conditional distribution of quantiles for the imaging task. During calibration, QUTCC generates uncertainty bounds by iteratively querying the network for upper and lower quantiles, progressively refining the bounds to obtain a tighter interval that captures the desired coverage. We evaluate our method on several denoising tasks as well as compressive MRI reconstruction. Our method successfully pinpoints hallucinations in image estimates and consistently achieves tighter uncertainty intervals than prior methods while maintaining the same statistical coverage.

Medical imaging plays a vital role in early disease diagnosis and monitoring. Specifically, blood microscopy offers valuable insights into blood cell morphology and the detection of hematological disorders. In recent years, deep learning-based automated classification systems have demonstrated high potential in enhancing the accuracy and efficiency of blood image analysis. However, a detailed performance analysis of specific deep learning frameworks appears to be lacking. This paper compares the performance of three popular deep learning frameworks, TensorFlow with Keras, PyTorch, and JAX, in classifying blood cell images from the publicly available BloodMNIST dataset. The study primarily focuses on inference time differences, but also classification performance for different image sizes. The results reveal variations in performance across frameworks, influenced by factors such as image resolution and framework-specific optimizations. Classification accuracy for JAX and PyTorch was comparable to current benchmarks, showcasing the efficiency of these frameworks for medical image classification.

Magnetic resonance imaging (MRI) has become a pivotal non-invasive tool in the evaluation and management of lymphedema. This review systematically summarizes its current applications, highlighting imaging techniques, comparative advantages over other modalities, MRI-based staging systems, and emerging clinical roles. A comprehensive literature review was conducted, covering comparisons with lymphoscintigraphy, ultrasound, and computed tomography (CT), as well as studies on the feasibility of multiparametric MRI sequences. Compared to conventional imaging, MRI offers superior soft tissue contrast and enables detailed assessment of lymphatic anatomy, tissue composition, and fluid distribution through sequences such as T2-weighted imaging, diffusion-weighted imaging (DWI), and magnetic resonance lymphangiography (MRL). Standardized grading systems have been proposed to support clinical staging. MRI is increasingly applied in preoperative planning and postoperative surveillance.These findings underscore MRI's diagnostic precision and clinical utility. Future research should focus on protocol standardization, incorporation of quantitative biomarkers, and development of AI-driven tools to enable personalized, scalable lymphedema care.

To assess the accuracy and time efficiency of manual versus artificial intelligence (AI)-driven tooth segmentation on cone-beam computed tomography (CBCT) images, using AI tools integrated within implant planning software, and to evaluate the impact of artifacts, dental arch, tooth type, and region. Fourteen patients who underwent CBCT scans were randomly selected for this study. Using the acquired datasets, 67 extracted teeth were segmented using one manual and two AI-driven tools. The segmentation time for each method was recorded. The extracted teeth were scanned with an intraoral scanner to serve as the reference. The virtual models generated by each segmentation method were superimposed with the surface scan models to calculate volumetric discrepancies. The discrepancy between the evaluated AI-driven and manual segmentation methods ranged from 0.10 to 0.98 mm, with a mean RMS of 0.27 (0.11) mm. Manual segmentation resulted in less RMS deviation compared to both AI-driven methods (CDX; BSB) (p < 0.05). Significant differences were observed between all investigated segmentation methods, both for the overall tooth area and each region, with the apical portion of the root showing the lowest accuracy (p < 0.05). Tooth type did not have a significant effect on segmentation (p > 0.05). Both AI-driven segmentation methods reduced segmentation time compared to manual segmentation (p < 0.05). AI-driven segmentation can generate reliable virtual 3D tooth models, with accuracy comparable to that of manual segmentation performed by experienced clinicians, while also significantly improving time efficiency. To further enhance accuracy in cases involving restoration artifacts, continued development and optimization of AI-driven tooth segmentation models are necessary.