Nature. 2025 Jul 9. doi: 10.1038/s41586-025-09205-6. Online ahead of print.

ABSTRACT

Advances in deep learning and AlphaFold2 have enabled the large-scale prediction of protein structures across species, opening avenues for studying protein function and evolution1. Here we analyse 11,269 predicted and experimentally determined enzyme structures that catalyse 361 metabolic reactions across 225 pathways to investigate metabolic evolution over 400 million years in the Saccharomycotina subphylum2. By linking sequence divergence in structurally conserved regions to a variety of metabolic properties of the enzymes, we reveal that metabolism shapes structural evolution across multiple scales, from species-wide metabolic specialization to network organization and the molecular properties of the enzymes. Although positively selected residues are distributed across various structural elements, enzyme evolution is constrained by reaction mechanisms, interactions with metal ions and inhibitors, metabolic flux variability and biosynthetic cost. Our findings uncover hierarchical patterns of structural evolution, in which structural context dictates amino acid substitution rates, with surface residues evolving most rapidly and small-molecule-binding sites evolving under selective constraints without cost optimization. By integrating structural biology with evolutionary genomics, we establish a model in which enzyme evolution is intrinsically governed by catalytic function and shaped by metabolic niche, network architecture, cost and molecular interactions.

PMID:40634610 | DOI:10.1038/s41586-025-09205-6

Sci Rep. 2025 Jul 9;15(1):24643. doi: 10.1038/s41598-025-10786-5.

ABSTRACT

Dissolved oxygen (DO) is a vital parameter in regulating water quality and sustaining the health of aquatic organisms in aquaculture environments. Therefore, estimation and control of DO levels are essential in aquaculture operations. However, traditional chemical and photochemical approaches are limited by inaccuracies, environmental interferences, time consumption, and the inability to provide real-time data. Recently, artificial intelligence techniques have been studied for DO estimation. However, off-the-shelf models, such as Random Forest (RF) and Back Propagation (BP) have demonstrated poor performance due to intricate interactions in aquatic ecosystems, which leads to complex data patterns. This study proposes a water quality estimation model by combining a convolutional neural network (CNN), self-attention (SA), and bidirectional simple recurrent unit (BiSRU). One-dimensional convolution in CNN was employed to extract effective features and input into the SA mechanism to assign weights and emphasise crucial information. The model's accuracy is improved by incorporating BiSRU. This model evaluated the DO levels of the intensive aquaculture base in Nansha, Guangzhou City, Guangdong Province, China. The proposed CNN-SA-BiSRU achieved MSE, MAE, RMSE, and R2 of 0.0022, 0.0341, 0.0471, and 0.9765, respectively. The results of the experiments showed that the proposed model had a high level of accuracy in estimating the outcomes with minimal fluctuations in estimation errors. Moreover, accuracy for short-term prediction was significantly improved, surpassing the performance of existing methods. The highly accurate results indicate the potential of the proposed methodology for DO-level monitoring in aquaculture and its usage in the fishery industry.

PMID:40634584 | DOI:10.1038/s41598-025-10786-5

Sci Rep. 2025 Jul 9;15(1):24748. doi: 10.1038/s41598-025-10567-0.

ABSTRACT

This paper proposes a new deep learning and machine learning model for detecting deception and suppression jamming in Ublox-M8T receivers operating under GNSS interference. This solution employs XGBoost for real-time classification of jamming signals, implemented on an STM32H743 microcontroller to ensure ultra-low latency, making it suitable for navigation in various environments. This work's key contribution is integrating a windowing mechanism for pre-saturation alerts and early activation of jamming detection which enhances system reliability by distinguishing between high-credibility and low-credibility GNSS data under static and dynamic jamming conditions. To validate the model, a series of experiments were conducted using a software-defined radio transmitter to simulate jamming scenarios. Genuine GNSS and jamming signals were collected under controlled conditions, and the data were pre-processed through feature normalization, correlation analysis, and feature selection based on importance in the mentioned systems. The XGBoost classifier, trained and tested on this processed dataset, achieved a detection rate of 99.97%, a precision of 99.94%, and a Matthews correlation coefficient of 0.9992, with an average prediction time of only 20 microseconds per sample in the implemented mode, making it an excellent choice for real-time systems. Additionally, the windowing mechanism enhances system performance by proactively initiating countermeasures before reaching saturation, ensuring continuous operation during high-intensity jamming attacks.

PMID:40634565 | DOI:10.1038/s41598-025-10567-0

Sci Rep. 2025 Jul 9;15(1):24654. doi: 10.1038/s41598-025-10086-y.

ABSTRACT

The regulation of indoor thermal comfort is a critical aspect of smart building design, significantly influencing energy efficiency and occupant well-being. Traditional comfort models, such as Fanger's equation and adaptive approaches, often fall short in capturing individual occupant preferences and the dynamic nature of indoor environmental conditions. To overcome these limitations, we introduce a Digital Twin-driven framework integrated with an advanced attention-based Long Short-Term Memory (LSTM) model specifically tailored for personalised thermal comfort prediction and intelligent HVAC control. The attention mechanism effectively focuses on critical temporal features, enhancing both predictive performance and interpretability. Next, the Digital Twin enables the real-time simulation of indoor environments and occupant responses, facilitating proactive comfort management. We utilise a subset of the ASHRAE Global Thermal Comfort Database II, and extensive pre-processing, including median-based data imputation and feature normalisation, is conducted. The proposed model categorises Thermal Sensation Votes (TSVs) recorded on a 7-point ASHRAE scale into three classes: Uncomfortably Cold (UC) for TSV ≤-1, Neutral (N) for TSV = 0, and Uncomfortably Warm (UW) for TSV ≥+1. The model achieves a test accuracy of 83.8%, surpassing previous state-of-the-art methods. Furthermore, Explainable AI (XAI) techniques, such as SHAP and LIME, are integrated to enhance transparency and interpretability, complemented by scenario-based energy efficiency analyses to evaluate energy-comfort trade-offs. This comprehensive approach provides a robust, interpretable, and energy-efficient solution for occupant-centric HVAC management in smart building systems.

PMID:40634515 | DOI:10.1038/s41598-025-10086-y

Sci Rep. 2025 Jul 9;15(1):24647. doi: 10.1038/s41598-025-09394-0.

ABSTRACT

Diabetic retinopathy (DR) is an age-related macular degeneration eye disease problem that causes pathological changes in the retinal neural and vascular system. Recently, fundus imaging is a popular technology and widely used for clinical diagnosis, diabetic retinopathy, etc. It is evident from the literature that image quality changes due to uneven illumination, pigmentation level effect, and camera sensitivity affect clinical performance, particularly in automated image analysis systems. In addition, low-quality retinal images make the subsequent precise segmentation a challenging task for the computer diagnosis of retinal images. Thus, in order to solve this issue, herein, we proposed an adaptive enhancement-based Deep Convolutional Neural Network (DCNN) model for diabetic retinopathy (DR). In our proposed model, we used an adaptive gamma enhancement matrix to optimize the color channels and contrast standardization used in images. The proposed model integrates quantile-based histogram equalization to expand the perceptibility of the fundus image. Our proposed model provides a remarkable improvement in fundus color images and can be used particularly for low-contrast quality images. We performed several experiments, and the efficiency is evaluated using a large public dataset named Messidor's. Our proposed model efficiently classifies a distinct group of retinal images. The average assessment score for the original and enhanced images is 0.1942 (standard deviation: 0.0799), Peak Signal-to-Noise Ratio (PSNR) 28.79, and Structural Similarity Index (SSIM) 0.71. The best classification accuracy is [Formula: see text], indicating that Convolutional Neural Networks (CNNs) and transfer learning are superior to traditional methods. The results show that the proposed model increases the contrast of a particular color image without altering its structural information.

PMID:40634513 | DOI:10.1038/s41598-025-09394-0

Clin Radiol. 2025 Jun 11;88:106987. doi: 10.1016/j.crad.2025.106987. Online ahead of print.

ABSTRACT

BACKGROUND: Conventional magnetic resonance imaging (MRI) protocols for lower back pain require multiple sequences and long acquisition times, challenging healthcare systems amid rising demand for lumbar spine imaging.

AIM: To assess the diagnostic performance of an abbreviated, deep learning-accelerated sagittal T2w Dixon single sequence protocol (Protocolabb-DL) versus the standard lumbar spine MRI protocol (Protocolstd).

MATERIALS AND METHODS: In this prospective, single-centre study, 30 patients (mean age: 48 ± 18.5 years; 67% female) with lower back pain (LBP) underwent a single MRI examination using both Protocolstd (sagittal T1w and T2w turbo spin-echo sequences) and Protocolabb-DL. A senior radiologist (15 years experience) established the diagnostic reference standard using Protocolstd. Two independent readers (10 and 5 years' experience) evaluated the images at a segmental level for degenerative pathologies, including Modic changes, disc pathology, facet arthropathy, neuroforaminal stenosis, and Schmorl nodes. Diagnostic performance, confidence, interprotocol, and interobserver agreements were analysed.

RESULTS: Protocolabb-DL reduced acquisition time by 80% at 1.5 Tesla (1:33 vs 7:43 minutes) and 84% at 3 Tesla (1:26 vs 8:43 minutes). Diagnostic performance was high, with sensitivities up to 100% [95% CI, 90.7-100.0] for Modic changes and 94.7% [95% CI, 87.1-98.5] for disc pathology, and specificities up to 100% [95% CI, 97.8-100.0] for Schmorl nodes. Diagnostic confidence was comparable between protocols (P > 0.05). Interprotocol agreement was excellent (κ: 0.84-1.00), and interobserver agreement for Protocolabb-DL was substantial to excellent (κ: 0.67-0.93).

CONCLUSION: Protocolabb-DL provides diagnostic performance comparable to Protocolstd for degenerative lumbar spine pathologies while reducing acquisition time by up to 84%.

PMID:40633138 | DOI:10.1016/j.crad.2025.106987

IEEE J Biomed Health Inform. 2025 Jul 9;PP. doi: 10.1109/JBHI.2025.3587103. Online ahead of print.

ABSTRACT

Multi-variate time-series are one of the primary data modalities involved in large classes of problems, where deep learning models represent the state-of-the-art solution. In the healthcare domain electrophysiological data, such as intracranial electroencephalography (iEEG), is used to perform a variety of tasks. However, iEEG models require that the number of channels be fixed, while iEEG setups in clinics are highly personalized and thus vary considerably from one subject to the next. To address this concern, we propose a channel-adaptive (CA) architecture that seamlessly functions on any multi-variate signal with an arbitrary number of channels. Each CA-model can be pre-trained on a large corpus of iEEG recordings from multiple heterogeneous subjects, and then finetuned to each subject using equal or lower amounts of data compared to existing state-of-the-art models, and in only 1/5 of the time. We evaluate our CA-models on a seizure detection task both on a short-term ($\sim$15 hours) and a long-term ($\sim$2600 hours) dataset. In particular, our CA-EEGWaveNet - based on EEGWaveNet - is trained on a single seizure of the tested subject, while the baseline EEGWaveNet is trained on all but one. CA-EEGWaveNet surpasses the baseline in median F1-score (0.78 vs 0.76). Similarly, CA-EEGNet - based on EEGNet - also surpasses its baseline (0.79 vs 0.74). Overall, we show that the CA architecture is a drop-in replacement for existing seizure classification models, bringing better characteristics and performance across the board.

PMID:40633043 | DOI:10.1109/JBHI.2025.3587103

IEEE Trans Neural Syst Rehabil Eng. 2025 Jul 9;PP. doi: 10.1109/TNSRE.2025.3587637. Online ahead of print.

ABSTRACT

Auditory attention decoding (AAD) is the process of identifying the attended speech in a multi-talker environment using brain signals, typically recorded through electroencephalography (EEG). Over the past decade, AAD has undergone continuous development, driven by its promising application in neuro-steered hearing devices. Most AAD algorithms are relying on the increase in neural entrainment to the envelope of attended speech, as compared to unattended speech, typically using a two-step approach. First, the algorithm predicts representations of the attended speech signal envelopes; second, it identifies the attended speech by finding the highest correlation between the predictions and the representations of the actual speech signals. In this study, we proposed a novel end-to-end neural network architecture, named AADNet, which combines these two stages into a direct approach to address the AAD problem. We compare the proposed network against traditional stimulus decoding-based approaches, including linear stimulus reconstruction, canonical correlation analysis, and an alternative non-linear stimulus reconstruction using three different datasets. AADNet shows a significant performance improvement for both subject-specific and subject-independent models. Notably, the average subject-independent classification accuracies for different analysis window lengths range from 56.3% (1 s) to 78.1% (20 s), 57.5% (1 s) to 89.4% (40 s), and 56.0% (1 s) to 82.6% (40 s) for three validated datasets, respectively, showing a significantly improved ability to generalize to data from unseen subjects. These results highlight the potential of deep learning models for advancing AAD, with promising implications for future hearing aids, assistive devices, and clinical assessments.

PMID:40633040 | DOI:10.1109/TNSRE.2025.3587637

J Chem Inf Model. 2025 Jul 9. doi: 10.1021/acs.jcim.5c00641. Online ahead of print.

ABSTRACT

In recent years, generative deep learning has emerged as a transformative approach in drug design, promising to explore the vast chemical space and generate novel molecules with desired biological properties. This perspective examines the challenges and opportunities of applying generative models to drug discovery, focusing on the intricate tasks related to small molecule generation, evaluation, and prioritization. Central to this process is navigating conflicting information from diverse sources─balancing chemical diversity, synthesizability, and bioactivity. We discuss the current state of generative methods, their optimization, and the critical need for robust evaluation protocols. By mapping this evolving landscape, we outline key building blocks, inherent dilemmas, and future directions in the journey to fully harness generative deep learning in the "chemical odyssey" of drug design.

PMID:40632942 | DOI:10.1021/acs.jcim.5c00641

Sci Adv. 2025 Jul 11;11(28):eadw4871. doi: 10.1126/sciadv.adw4871. Epub 2025 Jul 9.

ABSTRACT

Given the graft-versus-leukemia effect observed with allogeneic hematopoietic stem cell transplantation in refractory or relapsed acute myeloid leukemia (AML), immunotherapies have been explored in nontransplant settings. We applied a multiomic approach to examine bone marrow interactions in patients with AML treated with pembrolizumab and decitabine. Using extensively trained nuclear and membrane segmentation models, we achieved precise transcript assignment and deep learning-based image analysis. To address read-depth limitations, we integrated single-cell RNA sequencing with single-cell spatial transcriptomics from the same sample. Quantifying cell-cell distances at the edge level enabled more accurate tumor microenvironment analysis, revealing global and local immune cell enrichment near leukemia cells postpembrolizumab treatment, potentially linked to clinical response. Furthermore, ligand-receptor analysis indicated potential alterations in specific signaling pathways between leukemia and immune cells following immunotherapy treatment. These findings provide insights into immune interactions in AML and may inform therapeutic strategies.

PMID:40632867 | DOI:10.1126/sciadv.adw4871

Med Biol Eng Comput. 2025 Jul 9. doi: 10.1007/s11517-025-03388-w. Online ahead of print.

ABSTRACT

Parkinson's disease (PD) is a prevalent neurodegenerative disorder worldwide, often progressing to mild cognitive impairment (MCI) and dementia. Clinical diagnosis of PD mainly depends on characteristic motor symptoms, which can lead to misdiagnosis, underscoring the need for reliable biomarkers. Early detection of PD and effective monitoring of disease progression are crucial for enhancing patient outcomes. Electroencephalogram (EEG) signals, as non-invasive neural recordings, show great promise as diagnostic biomarkers. In this study, we present a novel approach for PD diagnosis through the analysis of EEG signals from distinct brain regions. We used two publicly available EEG datasets and constructed three-dimensional (3D) time-frequency spectrograms for each brain region using the continuous wavelet transform (CWT). To improve feature representation, these spectrograms were encoded in the red-green-blue (RGB) color space. A ResNet18 model was trained separately on the spectrograms of each brain region, and its performance was assessed using the leave-one-subject-out cross-validation (LOSOCV) method. The proposed method achieved classification accuracies of 92.86% and 90.32% on the two datasets, respectively. The experimental results confirm the efficacy of our approach, highlighting its potential as a valuable tool to aid clinical diagnosis of PD.

PMID:40632381 | DOI:10.1007/s11517-025-03388-w

Med Biol Eng Comput. 2025 Jul 9. doi: 10.1007/s11517-025-03406-x. Online ahead of print.

ABSTRACT

Reliable feedback of gait variables, such as joint moments, is critical for designing controllers of intelligent assistive devices that can assist the wearer outdoors. To estimate lower extremity joint moments quickly and accurately outside the laboratory, a novel multimodal motion intent recognition system by fusing traditional deep learning models is proposed in this paper. The developed estimation method uses the joint kinematics data and individual feature parameters to estimate lower limb joint moments in the sagittal plane under different motion conditions: walking, running, and stair ascent and descent. Specifically, seven deep learning models that use combination of convolutional neural network, recurrent neural networks and attention mechanisms as the unit models of the framework are designed. To improve the performance of the unit models, a data augmentation module is designed in the system. Using those unit models, a novel framework, DeepMPSF-Net, which treats the output of each unit model as a pseudo-sensor observation and utilizes variable weight fusion methods to improve classification accuracy and kinetics estimation performance, is proposed. The results show that the augmented DeepMPSF-Net can accurately identify the locomotion, and the estimation performance (PCC) of joint moments is improved to 0.952 (walking), 0.988 (running), 0.925 (stair ascent), and 0.921 (stair descent), respectively. It also suggests that the estimation system is expected to contribute to the development of intelligent assistive devices for the lower limbs.

PMID:40632380 | DOI:10.1007/s11517-025-03406-x

Cell Regen. 2025 Jul 9;14(1):30. doi: 10.1186/s13619-025-00251-z.

NO ABSTRACT

PMID:40632365 | DOI:10.1186/s13619-025-00251-z

J Med Syst. 2025 Jul 9;49(1):97. doi: 10.1007/s10916-025-02228-6.

ABSTRACT

Traditional cuffless blood pressure (BP) estimation methods often require collecting physiological signals, such as electrocardiogram (ECG) and photoplethysmography (PPG), from two distinct body sites to compute metrics like pulse transit time (PTT) or pulse arrival time (PAT). While these metrics strongly correlate with BP, their reliance on multiple signal sources and susceptibility to noise from modern wearable devices present significant challenges. Addressing these limitations, we propose an innovative framework that requires only PPG signals from a single body site, leveraging advancements in artificial intelligence and computer vision. Our approach employs images of PPG signals, along with their first (vPPG) and second (aPPG) derivatives, for enhanced BP estimation. ResNet-50 is utilized to extract features and identify regions within the PPG, vPPG, and aPPG images that correlate strongly with BP. These features are further refined using multi-head cross-attention (MHCA) mechanism, enabling efficient information exchange across the modalities derived from ResNet-50 outputs, thereby improving estimation accuracy. The framework is validated on three distinct datasets, demonstrating superior performance compared to traditional PAT and PTT-based methods. Furthermore, it adheres to stringent medical standards, such as those defined by the Association for the Advancement of Medical Instrumentation (AAMI) and the British Hypertension Society (BHS), ensuring clinical reliability. By reducing the need for multiple signal sources and incorporating cutting-edge AI techniques, this framework represents a significant advancement in non-invasive BP monitoring, offering a more practical and accurate alternative to traditional methodologies.

PMID:40632332 | DOI:10.1007/s10916-025-02228-6

J Cancer Res Clin Oncol. 2025 Jul 9;151(7):209. doi: 10.1007/s00432-025-06199-6.

ABSTRACT

BACKGROUND: Colon cancer remains a leading cause of cancer-related mortality globally, highlighting the urgent need for advanced diagnostic methods to improve early detection and patient outcomes.

METHODS: This study introduces ColoViT, a hybrid diagnostic framework that synergistically integrates EfficientNet and Vision Transformers. EfficientNet contributes scalability and high performance in feature extraction, while Vision Transformers effectively capture the global contextual information within colonoscopic images.

RESULTS: The integration of these models enables ColoViT to deliver precise and comprehensive image analysis, significantly improving the detection of precancerous lesions and early-stage colon cancers. The proposed model achieved a recall of 92.4%, precision of 98.9%, F1-score of 98.4%, and an AUC of 99% in our preliminary evaluation.

CONCLUSION: ColoViT demonstrates superior performance over existing models, offering a robust solution for enhancing the early detection of colon cancer through deep learning-based image analysis.

PMID:40632312 | DOI:10.1007/s00432-025-06199-6

Radiol Artif Intell. 2025 Jul 9:e240769. doi: 10.1148/ryai.240769. Online ahead of print.

ABSTRACT

"Just Accepted" papers have undergone full peer review and have been accepted for publication in Radiology: Artificial Intelligence. This article will undergo copyediting, layout, and proof review before it is published in its final version. Please note that during production of the final copyedited article, errors may be discovered which could affect the content. Purpose To improve the generalizability of pathologic complete response (pCR) prediction following neoadjuvant chemotherapy using deep learning (DL)-based retrospective pharmacokinetic quantification (RoQ) of early-treatment dynamic contrast-enhanced (DCE) MRI. Materials and Methods This multicenter retrospective study included breast MRI data from four publicly available datasets of patients with breast cancer acquired from May 2002 to November 2016. RoQ was performed using a previously developed DL model for clinical multiphasic DCE-MRI datasets. Radiomic analysis was performed on RoQ maps and conventional enhancement maps. These data, together with clinicopathologic variables and shape-based radiomic analysis, were subsequently applied in pCR prediction using logistic regression. Prediction performance was evaluated by area under the receiver operating characteristic curve (AUC). Results A total of 1073 female patients with breast cancer were included. The proposed method showed improved consistency and generalizability compared with the reference method, achieving higher AUCs across external datasets (0.82 [CI: 0.72-0.91], 0.75 [CI: 0.71-0.79], and 0.77 [CI: 0.66-0.86] for Datasets A2, B, and C, respectively). On Dataset A2 (from the same study as the training dataset), there was no significant difference in performance between the proposed method and reference method (P = .80). Notably, on the combined external datasets, the proposed method significantly outperformed the reference method (AUC: 0.75 [CI: 0.72- 0.79] vs 0.71 [CI: 0.68-0.76], P = .003). Conclusion This work offers a novel approach to improve the generalizability and predictive accuracy of pCR response in breast cancer across diverse datasets, achieving higher and more consistent AUC scores than existing methods. ©RSNA, 2025.

PMID:40631989 | DOI:10.1148/ryai.240769

Sports Biomech. 2025 Jul 9:1-16. doi: 10.1080/14763141.2025.2526702. Online ahead of print.

ABSTRACT

We used raw data from wearable sensors in consumer-realistic locations (replicating watch, arm phone strap, chest strap, etc.) to estimate lower-limb sagittal-plane joint moments during treadmill running and assessed the effect of a reduced number of sensor locations on estimation accuracy. Fifty mixed-ability runners (25 men and 25 women) ran on a treadmill at a range of speeds and gradients. Their data was used to train Long Short-Term Memory (LSTM) models in a supervised fashion. Estimation accuracy was evaluated by comparing model outputs against the criterion signals, calculated from marker-based kinematics and instrumented treadmill kinetics via inverse dynamics. The model that utilised data from all sensor locations achieved the lowest estimation error with a mean relative Root Mean Squared Error (rRMSE) of 12.1%, 9.0%, and 6.7% at the hip, knee, and ankle, respectively. Reducing data input to fewer sensors did not greatly compromise estimation accuracy. For example, a wrist-foot sensor combination only increased estimation error by 0.8% at the hip, and 1.0% at the knee and ankle joints. This work contributes to the development of a field-oriented tool that can provide runners with insight into their joint-level net moment contributions whilst leveraging data from their possible existing wearable sensor locations.

PMID:40631968 | DOI:10.1080/14763141.2025.2526702

Dermatol Surg. 2025 Jul 9. doi: 10.1097/DSS.0000000000004758. Online ahead of print.

ABSTRACT

BACKGROUND: There is a scarcity of artificial intelligence models trained on frozen pathology. One way to expand the clinical utility of models trained on permanent pathology is by applying them to frozen sections and fine-tune based on weaknesses.

OBJECTIVE: To qualitatively evaluate a deep learning model trained on permanent pathology to classify squamous cell carcinoma on Mohs surgery frozen sections to learn model shortcomings and inform retraining and fine-tuning.

MATERIALS AND METHODS: The authors trained a model for classification of tumor on 746 skin biopsy slides and tested it on 15 Mohs surgery frozen sections. The authors estimated performance metrics and compared the regions of interest generated by the model with the original H&E slides.

RESULTS: The model achieved an AUC-ROC of 0.985 and 0.796 for tumor classification in permanent pathology and in frozen sections, respectively. Regions of interest for frozen sections with scarce tumor areas were inaccurate, focusing on normal tissue for slides classified as false negative, or highlighting structures different from tumor (e.g., inflammation, muscle, and nerves) for slides classified as true positive.

CONCLUSION: Deep anatomical structures more commonly present in Mohs frozen pathology might represent data out-of-distribution for models trained on permanent pathology, potentially leading to inadequate model outputs.

PMID:40631753 | DOI:10.1097/DSS.0000000000004758

Microsc Res Tech. 2025 Jul 9. doi: 10.1002/jemt.70027. Online ahead of print.

ABSTRACT

Uterine cancer also referred to as endometrial cancer, which significantly impacts the female reproductive organs. The early diagnosis increases the survival rates and also prevents the progression of endometrial cancer thereby, the novel Transfer Learning based Convolution Neural Network with Adam Gold Rush Optimization (TL-CNN_AdGRO) is proposed to classify endometrial cancer using histopathological images. The histopathological image is fed to the preprocessing phase, which uses an Adaptive Weighted Mean Filter (AWMF). Next, the segmentation of endometrial cancer is utilized by the Directional Connectivity Network (DConn-Net). Following segmentation, feature mining is carried out, which includes Local Boundary Summation Pattern (LBSP) and Local Gaber Binary Pattern Histogram Sequence Features (LGBPHS). Finally, the endometrial cancer classification is achieved using TL-CNN by employing hyperparameters from the Xception model. Here TL-CNN is trained by AdGRO algorithm, which is the combination of Adam Optimizer and Gold Rush Optimization. Compared to existing models, the proposed model achieves superior performance with an accuracy of 91.876%, a True Positive Rate (TPR) of 93.987%, and a True Negative Rate (TNR) of 89.876% for K-sample 8. The results confirm the effectiveness of TL-CNN_AdGRO, also it demonstrates strong performance, ensures robustness, improves the early detection of endometrial cancer, and making it a promising approach for histopathological image analysis.

PMID:40631670 | DOI:10.1002/jemt.70027

J Med Imaging Radiat Oncol. 2025 Jul 9. doi: 10.1111/1754-9485.13880. Online ahead of print.

ABSTRACT

The integration of artificial intelligence (AI) into clinical practice, particularly within radiology, nuclear medicine and radiation oncology, is transforming diagnostic and therapeutic processes. AI-driven tools, especially in deep learning and machine learning, have shown remarkable potential in enhancing image recognition, analysis and decision-making. This technological advancement allows for the automation of routine tasks, improved diagnostic accuracy, and the reduction of human error, leading to more efficient workflows. Moreover, the successful implementation of AI in healthcare requires comprehensive education and training for young clinicians, with a pressing need to incorporate AI into residency programmes, ensuring that future specialists are equipped with traditional skills and a deep understanding of AI technologies and their clinical applications. This includes knowledge of software, data analysis, imaging informatics and ethical considerations surrounding AI use in medicine. By fostering interdisciplinary integration and emphasising AI education, healthcare professionals can fully harness AI's potential to improve patient outcomes and advance the field of medical imaging and therapy. This review aims to evaluate how AI influences radiology, nuclear medicine and radiation oncology, while highlighting the necessity for specialised AI training in medical education to ensure its successful clinical integration.

PMID:40631621 | DOI:10.1111/1754-9485.13880