Front Plant Sci. 2025 Feb 13;15:1491593. doi: 10.3389/fpls.2024.1491593. eCollection 2024.

ABSTRACT

A variety of diseased leaves and background noise types are present in images of diseased tomatoes captured in real-world environments. However, existing tomato leaf disease recognition models are limited to recognizing only a single leaf, rendering them unsuitable for practical applications in real-world scenarios. Additionally, these models consume significant hardware resources, making their implementation challenging for agricultural production and promotion. To address these issues, this study proposes a framework that integrates tomato leaf detection with leaf disease recognition. This framework includes a leaf detection model designed for diverse and complex environments, along with an ultra-lightweight model for recognizing tomato leaf diseases. To minimize hardware resource consumption, we developed five inverted residual modules coupled with an efficient attention mechanism, resulting in an ultra-lightweight recognition model that effectively balances model complexity and accuracy. The proposed network was trained on a dataset collected from real environments, and 14 contrasting experiments were conducted under varying noise conditions. The results indicate that the accuracy of the ultra-lightweight tomato disease recognition model, which utilizes the efficient attention mechanism, is 97.84%, with only 0.418 million parameters. Compared to traditional image recognition models, the model presented in this study not only achieves enhanced recognition accuracy across 14 noisy environments but also significantly reduces the number of required model parameters, thereby overcoming the limitation of existing models that can only recognize single disease images.

PMID:40017620 | PMC:PMC11865201 | DOI:10.3389/fpls.2024.1491593

Front Plant Sci. 2025 Feb 13;15:1448669. doi: 10.3389/fpls.2024.1448669. eCollection 2024.

ABSTRACT

UAV image acquisition and deep learning techniques have been widely used in field hydrological monitoring to meet the increasing data volume demand and refined quality. However, manual parameter training requires trial-and-error costs (T&E), and existing auto-trainings adapt to simple datasets and network structures, which is low practicality in unstructured environments, e.g., dry thermal valley environment (DTV). Therefore, this research combined a transfer learning (MTPI, maximum transfer potential index method) and an RL (the MTSA reinforcement learning, Multi-Thompson Sampling Algorithm) in dataset auto-augmentation and networks auto-training to reduce human experience and T&E. Firstly, to maximize the iteration speed and minimize the dataset consumption, the best iteration conditions (MTPI conditions) were derived with the improved MTPI method, which shows that subsequent iterations required only 2.30% dataset and 6.31% time cost. Then, the MTSA was improved under MTPI conditions (MTSA-MTPI) to auto-augmented datasets, and the results showed a 16.0% improvement in accuracy (human error) and a 20.9% reduction in standard error (T&E cost). Finally, the MTPI-MTSA was used for four networks auto-training (e.g., FCN, Seg-Net, U-Net, and Seg-Res-Net 50) and showed that the best Seg-Res-Net 50 gained 95.2% WPA (accuracy) and 90.9% WIoU. This study provided an effective auto-training method for complex vegetation information collection, which provides a reference for reducing the manual intervention of deep learning.

PMID:40017619 | PMC:PMC11864880 | DOI:10.3389/fpls.2024.1448669

Front Plant Sci. 2025 Feb 13;15:1474906. doi: 10.3389/fpls.2024.1474906. eCollection 2024.

ABSTRACT

Mung bean seeds are very important in agricultural production and food processing, but due to their variety and similar appearance, traditional classification methods are challenging, to address this problem this study proposes a deep learning-based approach. In this study, based on the deep learning model MobileNetV2, a DMS block is proposed for mung bean seeds, and by introducing the ECA block and Mish activation function, a high-precision network model, i.e., HPMobileNet, is proposed, which is explored to be applied in the field of image recognition for the fast and accurate classification of different varieties of mung bean seeds. In this study, eight different varieties of mung bean seeds were collected and a total of 34,890 images were obtained by threshold segmentation and image enhancement techniques. HPMobileNet was used as the main network model, and by training and fine-tuning on a large-scale mung bean seed image dataset, efficient feature extraction classification and recognition capabilities were achieved. The experimental results show that HPMobileNet exhibits excellent performance in the mung bean seed grain classification task, with the accuracy improving from 87.40% to 94.01% on the test set, and compared with other classical network models, the results show that HPMobileNet achieves the best results. In addition, this study analyzes the impact of the learning rate dynamic adjustment strategy on the model and explores the potential for further optimization and application in the future. Therefore, this study provides a useful reference and empirical basis for the development of mung bean seed classification and smart agriculture technology.

PMID:40017618 | PMC:PMC11865048 | DOI:10.3389/fpls.2024.1474906

J Cannabis Res. 2025 Feb 27;7(1):13. doi: 10.1186/s42238-025-00268-w.

ABSTRACT

BACKGROUND: Cannabis sativa L. with a rich history of traditional medicinal use, has garnered significant attention in contemporary research for its potential therapeutic applications in various human diseases, including pain, inflammation, cancer, and osteoarthritis. However, the specific molecular targets and mechanisms underlying the synergistic effects of its diverse phytochemical constituents remain elusive. Understanding these mechanisms is crucial for developing targeted, effective cannabis-based therapies.

METHODS: To investigate the molecular targets and pathways involved in the synergistic effects of cannabis compounds, we utilized DRIFT, a deep learning model that leverages attention-based neural networks to predict compound-target interactions. We considered both whole plant extracts and specific plant-based formulations. Predicted targets were then mapped to the Reactome pathway database to identify the biological processes affected. To facilitate the prediction of molecular targets and associated pathways for any user-specified cannabis formulation, we developed CANDI (Cannabis-derived compound Analysis and Network Discovery Interface), a web-based server. This platform offers a user-friendly interface for researchers and drug developers to explore the therapeutic potential of cannabis compounds.

RESULTS: Our analysis using DRIFT and CANDI successfully identified numerous molecular targets of cannabis compounds, many of which are involved in pathways relevant to pain, inflammation, cancer, and other diseases. The CANDI server enables researchers to predict the molecular targets and affected pathways for any specific cannabis formulation, providing valuable insights for developing targeted therapies.

CONCLUSIONS: By combining computational approaches with knowledge of traditional cannabis use, we have developed the CANDI server, a tool that allows us to harness the therapeutic potential of cannabis compounds for the effective treatment of various disorders. By bridging traditional pharmaceutical development with cannabis-based medicine, we propose a novel approach for botanical-based treatment modalities.

PMID:40016810 | DOI:10.1186/s42238-025-00268-w

Breast Cancer Res. 2025 Feb 27;27(1):30. doi: 10.1186/s13058-025-01971-5.

ABSTRACT

OBJECTIVE: The aim of this study was to develop and validate a deep learning radiomics (DLR) model based on longitudinal ultrasound data and clinical features to predict pathologic complete response (pCR) after neoadjuvant chemotherapy (NAC) in breast cancer patients.

METHODS: Between January 2018 and June 2023, 312 patients with histologically confirmed breast cancer were enrolled and randomly assigned to a training cohort (n = 219) and a test cohort (n = 93) in a 7:3 ratio. Next, pre-NAC and post-treatment 2-cycle ultrasound images were collected, and radiomics and deep learning features were extracted from NAC pre-treatment (Pre), post-treatment 2 cycle (Post), and Delta (pre-NAC-NAC 2 cycle) images. In the training cohort, to filter features, the intraclass correlation coefficient test, the Boruta algorithm, and the least absolute shrinkage and selection operator (LASSO) logistic regression were used. Single-modality models (Pre, Post, and Delta) were constructed based on five machine-learning classifiers. Finally, based on the classifier with the optimal predictive performance, the DLR model was constructed by combining Pre, Post, and Delta ultrasound features and was subsequently combined with clinical features to develop a combined model (Integrated). The discriminative power, predictive performance, and clinical utility of the models were further evaluated in the test cohort. Furthermore, patients were assigned into three subgroups, including the HR+/HER2-, HER2+, and TNBC subgroups, according to molecular typing to validate the predictability of the model across the different subgroups.

RESULTS: After feature screening, 16, 13, and 10 features were selected to construct the Pre model, Post model, and Delta model based on the five machine learning classifiers, respectively. The three single-modality models based on the XGBoost classifier displayed optimal predictive performance. Meanwhile, the DLR model (AUC of 0.827) was superior to the single-modality model (Pre, Post, and Delta AUCs of 0.726, 0.776, and 0.710, respectively) in terms of prediction performance. Moreover, multivariate logistic regression analysis identified Her-2 status and histological grade as independent risk factors for NAC response in breast cancer. In both the training and test cohorts, the Integrated model, which included Pre, Post, and Delta ultrasound features and clinical features, exhibited the highest predictive ability, with AUC values of 0.924 and 0.875, respectively. Likewise, the Integrated model displayed the highest predictive performance across the different subgroups.

CONCLUSION: The Integrated model, which incorporated pre-NAC treatment and early treatment ultrasound data and clinical features, accurately predicted pCR after NAC in breast cancer patients and provided valuable insights for personalized treatment strategies, allowing for timely adjustment of chemotherapy regimens.

PMID:40016785 | DOI:10.1186/s13058-025-01971-5

BMC Med. 2025 Feb 27;23(1):127. doi: 10.1186/s12916-025-03962-x.

ABSTRACT

BACKGROUND: Early diagnosis of biliary atresia (BA) is crucial for improving patient outcomes, yet remains a significant global challenge. This challenge may be ameliorated through the application of artificial intelligence (AI). Despite the promise of AI in medical diagnostics, its application to multimodal BA data has not yet achieved substantial breakthroughs. This study aims to leverage diverse data sources and formats to develop an intelligent diagnostic system for BA.

METHODS: We constructed the largest known multimodal BA dataset, comprising ultrasound images, clinical data, and laboratory results. Using this dataset, we developed a novel deep learning model and simplified it using easily obtainable data, eliminating the need for blood samples. The models were externally validated in a prospective study. We compared the performance of our model with human experts of varying expertise levels and evaluated the AI system's potential to enhance its diagnostic accuracy.

RESULTS: The retrospective study included 1579 participants. The multimodal model achieved an AUC of 0.9870 on the internal test set, outperforming human experts. The simplified model yielded an AUC of 0.9799. In the prospective study's external test set of 171 cases, the multimodal model achieved an AUC of 0.9740, comparable to that of a radiologist with over 10 years of experience (AUC = 0.9766). For less experienced radiologists, the AI-assisted diagnostic AUC improved from 0.6667 to 0.9006.

CONCLUSIONS: This AI-based screening application effectively facilitates early diagnosis of BA and serves as a valuable reference for addressing common challenges in rare diseases. The model's high accuracy and its ability to enhance the diagnostic performance of human experts underscore its potential for significant clinical impact.

PMID:40016769 | DOI:10.1186/s12916-025-03962-x

BMC Biol. 2025 Feb 27;23(1):63. doi: 10.1186/s12915-025-02166-2.

ABSTRACT

BACKGROUND: Cyclic peptides, known for their high binding affinity and low toxicity, show potential as innovative drugs for targeting "undruggable" proteins. However, their therapeutic efficacy is often hindered by poor membrane permeability. Over the past decade, the FDA has approved an average of one macrocyclic peptide drug per year, with romidepsin being the only one targeting an intracellular site. Biological experiments to measure permeability are time-consuming and labor-intensive. Rapid assessment of cyclic peptide permeability is crucial for their development.

RESULTS: In this work, we proposed a novel deep learning model, dubbed as MultiCycPermea, for predicting cyclic peptide permeability. MultiCycPermea extracts features from both the image information (2D structural information) and sequence information (1D structural information) of cyclic peptides. Additionally, we proposed a substructure-constrained feature alignment module to align the two types of features. MultiCycPermea has made a leap in predictive accuracy. In the in-distribution setting of the CycPeptMPDB dataset, MultiCycPermea reduced the mean squared error (MSE) by approximately 44.83% compared to the latest model Multi_CycGT (0.29 vs 0.16). By leveraging visual analysis tools, MultiCycPermea can reveal the relationship between peptide modification structures and membrane permeability, providing insights to improve the membrane permeability of cyclic peptides.

CONCLUSIONS: MultiCycPermea provides an effective tool that accurately predicts the permeability of cyclic peptides, offering valuable insights for improving the membrane permeability of cyclic peptides. This work paves a new path for the application of artificial intelligence in assisting the design of membrane-permeable cyclic peptides.

PMID:40016695 | DOI:10.1186/s12915-025-02166-2

BMC Bioinformatics. 2025 Feb 27;26(1):68. doi: 10.1186/s12859-025-06081-9.

ABSTRACT

BACKGROUND: Protein large language models (LLM) have been used to extract representations of enzyme sequences to predict their function, which is encoded by enzyme commission (EC) numbers. However, a comprehensive comparison of different LLMs for this task is still lacking, leaving questions about their relative performance. Moreover, protein sequence alignments (e.g. BLASTp or DIAMOND) are often combined with machine learning models to assign EC numbers from homologous enzymes, thus compensating for the shortcomings of these models' predictions. In this context, LLMs and sequence alignment methods have not been extensively compared as individual predictors, raising unaddressed questions about LLMs' performance and limitations relative to the alignment methods. In this study, we set out to assess the performance of ESM2, ESM1b, and ProtBERT language models in their ability to predict EC numbers, comparing them with BLASTp, against each other and against models that rely on one-hot encodings of amino acid sequences.

RESULTS: Our findings reveal that combining these LLMs with fully connected neural networks surpasses the performance of deep learning models that rely on one-hot encodings. Moreover, although BLASTp provided marginally better results overall, DL models provide results that complement BLASTp's, revealing that LLMs better predict certain EC numbers while BLASTp excels in predicting others. The ESM2 stood out as the best model among the LLMs tested, providing more accurate predictions on difficult annotation tasks and for enzymes without homologs.

CONCLUSIONS: Crucially, this study demonstrates that LLMs still have to be improved to become the gold standard tool over BLASTp in mainstream enzyme annotation routines. On the other hand, LLMs can provide good predictions for more difficult-to-annotate enzymes, particularly when the identity between the query sequence and the reference database falls below 25%. Our results reinforce the claim that BLASTp and LLM models complement each other and can be more effective when used together.

PMID:40016653 | DOI:10.1186/s12859-025-06081-9

BMC Cancer. 2025 Feb 27;25(1):367. doi: 10.1186/s12885-025-13740-w.

ABSTRACT

As reported by the International Agency for Research on Cancer (IARC), the global incidence of cancer reached nearly 20 million new cases in recent years, with cancer-related fatalities amounting to around 9.7 million. This underscores the profound impact cancer has on public health worldwide. Deep learning has become a mainstream approach in cancer recognition. Despite its significant progress, deep learning is known for its requirement of large quantities of labeled data. Few-shot learning addresses this limitation by reducing the need for extensive labeled samples. In the field of cancer recognition, data collection is particularly challenging due to the scarcity of categories compared to other fields, and current few-shot learning methods have not yielded satisfactory results. To tackle this, we propose an auxiliary meta-learning strategy for cancer recognition. During the auxiliary training phase, the feature mapping model is trained in conjunction with external data. This process neutralizes the prediction probability of misclassification, allowing the model to more readily learn distinguishing features and avoid performance degradation caused by discrepancies in external data. Additionally, the redundancy of some input principal components in the feature mapping model is reduced, while the implicit information within these components is extracted. The training process is further accelerated by utilizing depthwise over-parameterized convolutional layers. Moreover, the implementation of a three-branch structure contributes to faster training and enhanced performance. In the meta-training stage, the feature mapping model is optimized within the embedding space, utilizing category prototypes and cosine distance. During the meta-testing phase, a small number of labeled samples are employed to classify unknown data. We have conducted extensive experiments on the BreakHis, Pap smear, and ISIC 2018 datasets. The results demonstrate that our method achieves superior accuracy in cancer recognition. Furthermore, experiments on few-shot benchmark datasets indicate that our approach exhibits excellent generalization capabilities.

PMID:40016648 | DOI:10.1186/s12885-025-13740-w

Sci Rep. 2025 Feb 27;15(1):7009. doi: 10.1038/s41598-025-90646-4.

ABSTRACT

Detecting cassava leaf disease is challenging because it is hard to identify diseases accurately through visual inspection. Even trained agricultural experts may struggle to diagnose the disease correctly which leads to potential misjudgements. Traditional methods to diagnose these diseases are time-consuming, prone to error, and require expert knowledge, making automated solutions highly preferred. This paper explores the application of advanced deep learning techniques to detect as well as classify cassava leaf diseases which includes EfficientNet models, DenseNet169, Xception, MobileNetV2, ResNet models, Vgg19, InceptionV3, and InceptionResNetV2. A dataset consisting of around 36,000 labelled images of cassava leaves, afflicted by diseases such as Cassava Brown Streak Disease, Cassava Mosaic Disease, Cassava Green Mottle, Cassava Bacterial Blight, and healthy leaves, was used to train these models. Further the images were pre-processed by converting them into grayscale, reducing noise using Gaussian filter, obtaining the region of interest using Otsu binarization, Distance transformation, as well as Watershed technique followed by employing contour-based feature selection to enhance model performance. Models, after fine-tuned with ADAM optimizer computed that among the tested models, the hybrid model (DenseNet169 + EfficientNetB0) had superior performance with classification accuracy of 89.94% while as EfficientNetB0 had the highest values of precision, recall, and F1score with 0.78 each. The novelty of the hybrid model lies in its ability to combine DenseNet169's feature reuse capability with EfficientNetB0's computational efficiency, resulting in improved accuracy and scalability. These results highlight the potential of deep learning for accurate and scalable cassava leaf disease diagnosis, laying the foundation for automated plant disease monitoring systems.

PMID:40016508 | DOI:10.1038/s41598-025-90646-4

NPJ Digit Med. 2025 Feb 27;8(1):126. doi: 10.1038/s41746-025-01455-y.

ABSTRACT

Recognizing the limitations of computer-assisted tools for thyroid nodule diagnosis using static ultrasound images, this study developed a diagnostic tool utilizing dynamic ultrasound video, namely Thyroid Nodules Visualization (TNVis), by leveraging a two-stage deep learning framework that involved three-dimensional (3D) visualization. In this multicenter study, 4569 cases were included for framework development, and data from seven hospitals were employed for diagnostic validation. TNVis achieved a Dice similarity coefficient of 0.90 after internal testing. For the external validation, TNVis significantly improved radiologists' performance, reaching an AUC of 0.79, compared to their diagnostic performance without the use of TNVis (AUC: 0.66; p < 0.001) and those with partial assistance (AUC: 0.72; p < 0.001). In conclusion, the TNVis-assisted diagnostic strategy not only significantly improves the diagnostic ability of radiologists but also closely imitates their clinical diagnostic procedures and provides them with an objective 3D representation of the nodules for precise and personalized diagnosis and treatment planning.

PMID:40016505 | DOI:10.1038/s41746-025-01455-y

NPJ Digit Med. 2025 Feb 27;8(1):130. doi: 10.1038/s41746-025-01473-w.

ABSTRACT

Glaucoma, a leading cause of blindness, requires accurate early detection. We present an AI-based Glaucoma Screening (AI-GS) network comprising six lightweight deep learning models (total size: 110 MB) that analyze fundus images to identify early structural signs such as optic disc cupping, hemorrhages, and nerve fiber layer defects. The segmentation of the optic cup and disc closely matches that of expert ophthalmologists. AI-GS achieved a sensitivity of 0.9352 (95% CI 0.9277-0.9435) at 95% specificity. In real-world testing, sensitivity dropped to 0.5652 (95% CI 0.5218-0.6058) at ~0.9376 specificity (95% CI 0.9174-0.9562) for the standalone binary glaucoma classification model, whereas the full AI-GS network maintained higher sensitivity (0.8053, 95% CI 0.7704-0.8382) with good specificity (0.9112, 95% CI 0.8887-0.9356). The sub-models in AI-GS, with enhanced capabilities in detecting early glaucoma-related structural changes, drive these improvements. With low computational demands and tunable detection parameters, AI-GS promises widespread glaucoma screening, portable device integration, and improved understanding of disease progression.

PMID:40016437 | DOI:10.1038/s41746-025-01473-w

Sci Rep. 2025 Feb 27;15(1):7010. doi: 10.1038/s41598-025-92020-w.

ABSTRACT

Health is fundamental to human well-being, with brain health particularly critical for cognitive functions. Magnetic resonance imaging (MRI) serves as a cornerstone in diagnosing brain health issues, providing essential data for healthcare decisions. These images represent vast datasets that are increasingly harnessed by deep learning for high-performance image processing and classification tasks. In our study, we focus on classifying brain tumors-such as glioma, meningioma, and pituitary tumors-using the U-Net architecture applied to MRI scans. Additionally, we explore the effectiveness of convolutional neural networks including Inception-V3, EfficientNetB4, and VGG19, augmented through transfer learning techniques. Evaluation metrics such as F-score, recall, precision, and accuracy were employed to assess model performance. The U-Net segmentation architecture, emerged as the top performer, achieving an accuracy of 98.56%, along with an F-score of 99%, an area under the curve of 99.8%, and recall and precision rates of 99%. This study demonstrates the effectiveness of U-Net, a convolutional neural network architecture, for accurate brain tumor segmentation in early detection and treatment planning. Achieving an accuracy of 96.01% in cross-dataset validation with an external cohort, U-Net exhibited robust performance across diverse clinical scenarios. Our findings highlight the potential of U-Net and transfer learning in enhancing diagnostic accuracy and informing clinical decision-making in neuroimaging, ultimately improving patient care and outcomes.

PMID:40016334 | DOI:10.1038/s41598-025-92020-w

Proc Natl Acad Sci U S A. 2025 Mar 4;122(9):e2418643122. doi: 10.1073/pnas.2418643122. Epub 2025 Feb 27.

ABSTRACT

Single-particle tracking has enabled quantitative studies of complex systems, providing nanometer localization precision and millisecond temporal resolution in heterogeneous environments. However, at micro- or nanometer scales, probe dynamics become inherently stochastic due to Brownian motion and complex interactions, leading to varied diffusion behaviors. Typically, analysis of such trajectory data involves certain moving-window operation and assumes the existence of some pseudo-steady states, particularly when evaluating predefined parameters or specific types of diffusion modes. Here, we introduce the stochastic particle-informed neural network (SPINN), a physics-informed deep learning framework that integrates stochastic differential equations to model and infer particle diffusion dynamics. The SPINN autonomously explores parameter spaces and distinguishes between deterministic and stochastic components with single-frame resolution. Using the anomalous diffusion dataset, we validated SPINN's ability to reduce frame-to-frame variability while preserving key statistical correlations, allowing for accurate characterization of different stochastic processes. When applied to the diffusion of single gold nanorods in hydrogels, the SPINN revealed enhanced microrheological properties during hydrogel gelation and uncovered interfacial dynamics during dextran/tetra-PEG liquid-liquid phase separation. By improving the temporal resolution of stochastic dynamics, the SPINN facilitates the estimation and prediction of complex diffusion behaviors, offering insights into underlying physical mechanisms at mesoscopic scales.

PMID:40014572 | DOI:10.1073/pnas.2418643122

Microsc Res Tech. 2025 Feb 27. doi: 10.1002/jemt.24836. Online ahead of print.

ABSTRACT

The retina is an important part of the eye that aids in focusing light and visual recognition to the brain. Hence, its damage causes vision loss in the human eye. Central serous retinopathy is a common retinal disorder in which serous detachment occurs at the posterior pole of the retina. Therefore, detection of CSR at an early stage with good accuracy can decrease the rate of vision loss and recover the vision to normal conditions. In the past, numerous manual techniques have been devised for CSR detection; nevertheless, they have demonstrated imprecision and unreliability. Thus, the deep learning method can play an important role in automatically detecting CSR. This research presents a convolutional neural network-based framework combined with segmentation and post-ocessing for CSR classification. There are several challenges in the segmentation of retinal images, such as noise, size variation, location, and shape of the fluid in the retina. To address these limitations, Otsu's thresholding has been employed as a technique for segmenting optical coherence tomography (OCT) images. Pigments and fluids are present in epithelial detachment, and contrast adjustment and noise removal are required. After segmentation, post-processing is used, combining flood filling, dilation, and area thresholding. The segmented processed OCT scans were classified using the fusion of three networks: (i) ResNet-18, (ii) Google-Net, and (iii) VGG-19. After experimentation, the fusion of ResNet-18, GoogleNet, and VGG-19 achieved 99.6% accuracy, 99.46% sensitivity, 100% specificity, and 99.73% F1 score using the proposed framework for classifying normal and CSR-affected images. A publicly available dataset OCTID comprises 207 normal and 102 CSR-affected images was utilized for testing and training of the proposed method. The experimental findings conclusively demonstrate the inherent suitability and efficacy of the framework put forth. Through rigorous testing and analysis, the results unequivocally validate the framework's ability to fulfill its intended objectives and address the challenges at hand.

PMID:40014549 | DOI:10.1002/jemt.24836

Magn Reson Med. 2025 Feb 27. doi: 10.1002/mrm.30478. Online ahead of print.

ABSTRACT

PURPOSE: The clinical value of quantitative MRI hinges on its measurement repeatability. Deep learning methods to reconstruct undersampled quantitative MRI can accelerate reconstruction but do not aim to promote quantitative repeatability. This study proposes a repeatability-encouraging self-supervised learning (SSL) reconstruction method for quantitative MRI.

METHODS: The proposed SSL reconstruction network minimized cross-data-consistency between two equally sized, mutually exclusive temporal subsets of k-t-space data, encouraging repeatability by enabling each subset's reconstruction to predict the other's k-t-space data. The method was evaluated on cardiac MR Multitasking T1 mapping data and compared with supervised learning methods trained on full 60-s inputs (Sup60) and on split 30-s inputs (Sup30/30). Reconstruction quality was evaluated on full 60-s inputs, comparing results to iterative wavelet-regularized references using Bland-Altman limits of agreement (LOAs). Repeatability was evaluated by splitting the 60-s data into two 30-s inputs, evaluating T1 differences between reconstructions from the two halves of the scan.

RESULTS: On 60-s inputs, the proposed method produced comparable-quality images and T1 maps to the Sup60 method, with T1 values in general agreement with the wavelet reference (LOA Sup60 = ±75 ms, SSL = ±81 ms), whereas the Sup30/30 method generated blurrier results and showed poor T1 agreement (LOA Sup30/30 = ±132 ms). On back-to-back 30-s inputs, the proposed method had the best T1 repeatability (coefficient of variation SSL = 6.3%, Sup60 = 12.0%, Sup30/30 = 6.9%). Of the three deep learning methods, only the SSL method produced sharp and repeatable images.

CONCLUSION: Without the need for labeled training data, the proposed SSL method demonstrated superior repeatability compared with supervised learning without sacrificing sharpness, and reduced reconstruction time versus iterative methods.

PMID:40014485 | DOI:10.1002/mrm.30478

Transl Vis Sci Technol. 2025 Feb 3;14(2):26. doi: 10.1167/tvst.14.2.26.

ABSTRACT

PURPOSE: This study aims to develop an automated pipeline to detect retinal detachment from B-scan ocular ultrasonography (USG) images by using deep learning-based segmentation.

METHODS: A computational pipeline consisting of an encoder-decoder segmentation network and a machine learning classifier was developed, trained, and validated using 279 B-scan ocular USG images from 204 patients, including 66 retinal detachment (RD) images, 36 posterior vitreous detachment images, and 177 healthy control images. Performance metrics, including the precision, recall, and F-scores, were calculated for both segmentation and RD detection.

RESULTS: The overall pipeline achieved 96.3% F-score for RD detection, outperforming end-to-end deep learning classification models (ResNet-50 and MobileNetV3) with 94.3% and 95.0% F-scores. This improvement was also validated on an independent test set, where the proposed pipeline led to 96.5% F-score, but the classification models yielded only 62.1% and 84.9% F-scores, respectively. Besides, the segmentation model of this pipeline led to high performances across multiple ocular structures, with 84.7%, 78.3%, and 88.2% F-scores for retina/choroid, sclera, and optic nerve sheath segmentation, respectively. The segmentation model outperforms the standard UNet, particularly in challenging RD cases, where it effectively segmented detached retina regions.

CONCLUSIONS: The proposed automated segmentation and classification method improves RD detection in B-scan ocular USG images compared to end-to-end classification models, offering potential clinical benefits in resource-limited settings.

TRANSLATIONAL RELEVANCE: We have developed a novel deep/machine learning based pipeline that has the potential to significantly improve diagnostic accuracy and accessibility for ocular USG.

PMID:40014336 | DOI:10.1167/tvst.14.2.26

Eur J Nucl Med Mol Imaging. 2025 Feb 27. doi: 10.1007/s00259-025-07149-7. Online ahead of print.

ABSTRACT

PURPOSE: PET/CT imaging data contains a wealth of quantitative information that can provide valuable contributions to characterising tumours. A growing body of work focuses on the use of deep-learning (DL) techniques for denoising PET data. These models are clinically evaluated prior to use, however, quantitative image assessment provides potential for further evaluation. This work uses radiomic features to compare two manufacturer deep-learning (DL) image enhancement algorithms, one of which has been commercialised, against 'gold-standard' image reconstruction techniques in phantom data and a sarcoma patient data set (N=20).

METHODS: All studies in the retrospective sarcoma clinical [ 18 F]FDG dataset were acquired on either a GE Discovery 690 or 710 PET/CT scanner with volumes segmented by an experienced nuclear medicine radiologist. The modular heterogeneous imaging phantom used in this work was filled with [ 18 F]FDG, and five repeat acquisitions of the phantom were acquired on a GE Discovery 710 PET/CT scanner. The DL-enhanced images were compared to 'gold-standard' images the algorithms were trained to emulate and input images. The difference between image sets was tested for significance in 93 international biomarker standardisation initiative (IBSI) standardised radiomic features.

RESULTS: Comparing DL-enhanced images to the 'gold-standard', 4.0% and 9.7% radiomic features measured significantly different (pcritical < 0.0005) in the phantom and patient data respectively (averaged over the two DL algorithms). Larger differences were observed comparing DL-enhanced images to algorithm input images with 29.8% and 43.0% of radiomic features measuring significantly different in the phantom and patient data respectively (averaged over the two DL algorithms).

CONCLUSION: DL-enhanced images were found to be similar to images generated using the 'gold-standard' target image reconstruction method with more than 80% of radiomic features not significantly different in all comparisons across unseen phantom and sarcoma patient data. This result offers insight into the performance of the DL algorithms, and demonstrate potential applications for DL algorithms in harmonisation for radiomics and for radiomic features in quantitative evaluation of DL algorithms.

PMID:40014074 | DOI:10.1007/s00259-025-07149-7

J Appl Clin Med Phys. 2025 Feb 27:e70034. doi: 10.1002/acm2.70034. Online ahead of print.

ABSTRACT

Artificial intelligence (AI) has the potential to revolutionize brachytherapy's clinical workflow. This review comprehensively examines the application of AI, focusing on machine learning and deep learning, in various aspects of brachytherapy. We analyze AI's role in making brachytherapy treatments more personalized, efficient, and effective. The applications are systematically categorized into seven categories: imaging, preplanning, treatment planning, applicator reconstruction, quality assurance, outcome prediction, and real-time monitoring. Each major category is further subdivided based on cancer type or specific tasks, with detailed summaries of models, data sizes, and results presented in corresponding tables. Additionally, we discuss the limitations, challenges, and ethical concerns of current AI applications, along with perspectives on future directions. This review offers insights into the current advancements, challenges, and the impact of AI on treatment paradigms, encouraging further research to expand its clinical utility.

PMID:40014044 | DOI:10.1002/acm2.70034

J Neurophysiol. 2025 Feb 27. doi: 10.1152/jn.00622.2024. Online ahead of print.

ABSTRACT

Dendritic spines, small protrusions on neuronal dendrites, play a crucial role in brain function by changing shape and size in response to neural activity. So far, in depth analysis of dendritic spines in human brain tissue is lacking. This study presents a comprehensive analysis of human dendritic spine morphology and density using a unique dataset from human brain tissue from 27 patients (8 females, 19 males, aged 18-71) undergoing tumor or epilepsy surgery at three neurosurgery sites. We used acute slices and organotypic brain slice cultures to examine dendritic spines, classifying them into the three main morphological subtypes: Mushroom, Thin, and Stubby, via 3D reconstruction using ZEISS arivis Pro software. A deep learning model, trained on 39 diverse datasets, automated spine segmentation and 3D reconstruction, achieving a 74% F1-score and reducing processing time by over 50%. We show significant differences in spine density by sex, dendrite type, and tissue condition. Females had higher spine densities than males, and apical dendrites were denser in spines than basal ones. Acute tissue showed higher spine densities compared to cultured human brain tissue. With time in culture, Mushroom spines decreased, while Stubby and Thin spine percentages increased, particularly from 7-9 to 14 days in vitro, reflecting potential synaptic plasticity changes. Our study underscores the importance of using human brain tissue to understand unique synaptic properties and shows that integrating deep learning with traditional methods enables efficient large-scale analysis, revealing key insights into sex- and tissue-specific dendritic spine dynamics relevant to neurological diseases.

PMID:40013734 | DOI:10.1152/jn.00622.2024