Front Neurosci. 2024 Dec 10;18:1519970. doi: 10.3389/fnins.2024.1519970. eCollection 2024.

ABSTRACT

INTRODUCTION: Emotion recognition using electroencephalography (EEG) is a key aspect of brain-computer interface research. Achieving precision requires effectively extracting and integrating both spatial and temporal features. However, many studies focus on a single dimension, neglecting the interplay and complementarity of multi-feature information, and the importance of fully integrating spatial and temporal dynamics to enhance performance.

METHODS: We propose the Spatiotemporal Adaptive Fusion Network (STAFNet), a novel framework combining adaptive graph convolution and temporal transformers to enhance the accuracy and robustness of EEG-based emotion recognition. The model includes an adaptive graph convolutional module to capture brain connectivity patterns through spatial dynamic evolution and a multi-structured transformer fusion module to integrate latent correlations between spatial and temporal features for emotion classification.

RESULTS: Extensive experiments were conducted on the SEED and SEED-IV datasets to evaluate the performance of STAFNet. The model achieved accuracies of 97.89% and 93.64%, respectively, outperforming state-of-the-art methods. Interpretability analyses, including confusion matrices and t-SNE visualizations, were employed to examine the influence of different emotions on the model's recognition performance. Furthermore, an investigation of varying GCN layer depths demonstrated that STAFNet effectively mitigates the over-smoothing issue in deeper GCN architectures.

DISCUSSION: In summary, the findings validate the effectiveness of STAFNet in EEG-based emotion recognition. The results emphasize the critical role of spatiotemporal feature extraction and introduce an innovative framework for feature fusion, advancing the state of the art in emotion recognition.

PMID:39720230 | PMC:PMC11666491 | DOI:10.3389/fnins.2024.1519970

NPJ Digit Med. 2024 Dec 24;7(1):378. doi: 10.1038/s41746-024-01397-x.

ABSTRACT

Deep learning (DL) holds great promise to improve medical diagnostics, including pathology. Current DL research mainly focuses on performance. DL implementation potentially leads to environmental consequences but approaches for assessment of both performance and carbon footprint are missing. Here, we explored an approach for developing DL for pathology, which considers both diagnostic performance and carbon footprint, calculated as CO2 or equivalent emissions (CO2eq). We evaluated various DL architectures used in computational pathology, including a large foundation model, across two diagnostic tasks of low and high complexity. We proposed a metric termed 'environmentally sustainable performance' (ESPer), which quantitatively integrates performance and operational CO2eq during training and inference. While some DL models showed comparable diagnostic performance, ESPer enabled prioritizing those with less carbon footprint. We also investigated how data reduction approaches can improve the ESPer of individual models. This study provides an approach facilitating the development of environmentally friendly, sustainable medical AI.

PMID:39719527 | DOI:10.1038/s41746-024-01397-x

Sci Rep. 2024 Dec 24;14(1):30634. doi: 10.1038/s41598-024-82710-2.

ABSTRACT

Being able to quickly and accurately detect and recognize target balls is a key task for basketball robots in automated and intelligent sports competitions. This study aims to propose an effective target detection method for basketball robots, which is based on the improved YOLOv5s model and introduces spatial pyramid pooling and instance-batch normalization structure. The study first pre-trained the model and compared the migration training with the random initialization approach. The experimental outcomes denote that the migration-trained model obtains a mean average precision value of 0.918 on the target detection task, which is significantly better than the model trained from scratch. Then, the study applies different improvement schemes to the YOLOv5s model and compares the improvement effects of the various schemes. The experimental outcomes denote that scheme 2 has the best improvement effect on the YOLOv5s model, and its detection accuracy on dataset 1 is 94.5%. The experiment proves that the target detection algorithm designed in the study is effective and accurate, and can help the basketball robot successfully accomplish the target detection task. This research helps to advance the development of basketball robotics and provides theoretical support and technical basis for efficient automated basketball games in the future.

PMID:39719525 | DOI:10.1038/s41598-024-82710-2

Eur Arch Otorhinolaryngol. 2024 Dec 24. doi: 10.1007/s00405-024-09169-9. Online ahead of print.

ABSTRACT

BACKGROUND: Artificial intelligence (AI) demonstrates high potential when applied to radiomic analysis of magnetic resonance imaging (MRI) to discriminate sinonasal tumors. This can enhance diagnostic suspicion beyond visual assessment alone and prior to biopsy, leading to expedite the diagnostic timeline and the treatment planning. The aim of the present work is to evaluate the current advancements and accuracy of this technology in this domain.

METHODS: A systematic literature review was conducted following PRISMA guidelines. Inclusion criteria comprised studies utilizing any machine learning approach applied to MRI of patients with sinonasal tumors. For each study, comprehensive data were gathered on the MRI protocols, feature extraction techniques, and classifiers employed to develop the AI model. The performance was assessed based on accuracy and area under the curve (AUC).

RESULTS: Fourteen studies, published between May 2017 and August 2024, were included. These studies were categorized into three groups: those examining both benign and malignant tumors, those investigating malignant tumor subpopulations, and those focusing on benign pathologies. All studies reported an AUC greater than 0.800, achieving AUC > 0.89 and accuracy > 0.81 when incorporating clinical-radiological variables. Notably, the best discrimination performance was observed in studies utilizing combined conventional MRI sequences, including T1-weighted, contrasted T1-weighted, and T2-weighted images.

CONCLUSION: The application of AI and radiomics in analyzing MRI scans presents significant promise for improving the discrimination of sinonasal tumors. Integrating clinical and radiological indicators enhances model performance, suggesting that future research should focus on larger patient cohorts and diverse AI methodologies to refine diagnostic accuracy and clinical utility.

PMID:39719474 | DOI:10.1007/s00405-024-09169-9

Sci Rep. 2024 Dec 24;14(1):30633. doi: 10.1038/s41598-024-82624-z.

ABSTRACT

Recently, social demands for a good quality of life have increased among the elderly and disabled people. So, biomedical engineers and robotic researchers aimed to fuse these techniques in a novel rehabilitation system. Moreover, these models utilized the biomedical signals acquired from the human body's particular organ, cells, or tissues. The human motion intention prediction mechanism plays an essential role in various applications, such as assistive and rehabilitation robots, that execute specific tasks among elders and physically impaired individuals. However, more complications are increased in the human-machine-based interaction techniques, creating more scope for personalized assistance for the human motion intention prediction system. Therefore, in this paper, an Adaptive Hybrid Network (AHN) is implemented for effective human motion intention prediction. Initially, multimodal data like electroencephalogram (EEG)/Electromyography (EMG) signals and sensor measures data are collected from the available data resource. The gathered EEG/EMG signals are then converted into spectrogram images and sent to AH-CNN-LSTM, which is the integration of an Adaptive Hybrid Convolution Neural Network (AH-CNN) with a Long Short-Term Memory (LSTM) network. Similarly, the data details of sensor measures are directly subjected to AH-CNN-Res-LSTM, which is the combination of Adaptive Hybrid CNN with Residual Network and LSTM (Res-LSTM) to get the predictive result. Further, to enhance the prediction, the parameters in both the AH-CNN-LSTM and AH-CNN-Res-LSTM techniques are optimized using the Improved Yellow Saddle Goatfish Algorithm (IYSGA). The efficiency of the implemented model is computed by conducting the comparison experiment of the proposed technique with other standard models. The performance outcome of the developed method outperformed the other traditional methods.

PMID:39719464 | DOI:10.1038/s41598-024-82624-z

Microsyst Nanoeng. 2024 Dec 24;10(1):201. doi: 10.1038/s41378-024-00845-8.

ABSTRACT

Microscopic imaging is a critical tool in scientific research, biomedical studies, and engineering applications, with an urgent need for system miniaturization and rapid, precision autofocus techniques. However, traditional microscopes and autofocus methods face hardware limitations and slow software speeds in achieving this goal. In response, this paper proposes the implementation of an adaptive Liquid Lens Microscope System utilizing Deep Reinforcement Learning-based Autofocus (DRLAF). The proposed study employs a custom-made liquid lens with a rapid zoom response, which is treated as an "agent." Raw images are utilized as the "state", with voltage adjustments representing the "actions." Deep reinforcement learning is employed to learn the focusing strategy directly from captured images, achieving end-to-end autofocus. In contrast to methodologies that rely exclusively on sharpness assessment as a model's labels or inputs, our approach involved the development of a targeted reward function, which has proven to markedly enhance the performance in microscope autofocus tasks. We explored various action group design methods and improved the microscope autofocus speed to an average of 3.15 time steps. Additionally, parallel "state" dataset lists with random sampling training are proposed which enhances the model's adaptability to unknown samples, thereby improving its generalization capability. The experimental results demonstrate that the proposed liquid lens microscope with DRLAF exhibits high robustness, achieving a 79% increase in speed compared to traditional search algorithms, a 97.2% success rate, and enhanced generalization compared to other deep learning methods.

PMID:39719441 | DOI:10.1038/s41378-024-00845-8

RMD Open. 2024 Dec 23;10(4):e004628. doi: 10.1136/rmdopen-2024-004628.

ABSTRACT

PURPOSE: To examine whether incorporating anatomy-centred deep learning can improve generalisability and enable prediction of disease progression.

METHODS: This retrospective multicentre study included conventional pelvic radiographs of four different patient cohorts focusing on axial spondyloarthritis collected at university and community hospitals. The first cohort, which consisted of 1483 radiographs, was split into training (n=1261) and validation (n=222) sets. The other cohorts comprising 436, 340 and 163 patients, respectively, were used as independent test datasets. For the second cohort, follow-up data of 311 patients was used to examine progression prediction capabilities. Two neural networks were trained, one on images cropped to the bounding box of the sacroiliac joints (anatomy-centred) and the other one on full radiographs. The performance of the models was compared using the area under the receiver operating characteristic curve (AUC), accuracy, sensitivity and specificity.

RESULTS: On the three test datasets, the standard model achieved AUC scores of 0.853, 0.817, 0.947, with an accuracy of 0.770, 0.724, 0.850. Whereas the anatomy-centred model achieved AUC scores of 0.899, 0.846, 0.957, with an accuracy of 0.821, 0.744, 0.906, respectively. The patients who were identified as high risk by the anatomy-centred model had an OR of 2.16 (95% CI 1.19, 3.86) for having progression of radiographic sacroiliitis within 2 years.

CONCLUSION: Anatomy-centred deep learning can improve the generalisability of models in detecting radiographic sacroiliitis. The model is published as fully open source alongside this study.

PMID:39719299 | DOI:10.1136/rmdopen-2024-004628

Brain Res. 2024 Dec 22:149423. doi: 10.1016/j.brainres.2024.149423. Online ahead of print.

ABSTRACT

OBJECTIVES: This narrative review aims to analyze mechanisms underlying Brain-Computer Interface (BCI) and Artificial Intelligence (AI) integration, evaluate recent advances in signal acquisition and processing techniques, and assess AI-enhanced neural decoding strategies. The review identifies critical research gaps and examines emerging solutions across multiple domains of BCI-AI integration.

METHODS: A narrative review was conducted using major biomedical and scientific databases including PubMed, Web of Science, IEEE Xplore, and Scopus (2014-2024). Literature was analyzed to identify key developments in BCI-AI integration, with particular emphasis on recent advances (2019-2024). The review process involved thematic analysis of selected publications focusing on practical applications, technical innovations, and emerging challenges.

RESULTS: Recent advances demonstrate significant improvements in BCI-AI systems: 1) High-density electrode arrays achieve spatial resolution up to 5 mm, with stable recordings over 15 months; 2) Deep learning decoders show 40 % improvement in information transfer rates compared to traditional methods; 3) Adaptive algorithms maintain >90 % success rates in motor control tasks over 200-day periods without recalibration; 4) Novel closed-loop optimization frameworks reduce user training time by 55 % while improving accuracy. Latest developments in flexible neural interfaces and self-supervised learning approaches show promise in addressing long-term stability and cross-user generalization challenges.

CONCLUSIONS: BCI-AI integration shows remarkable progress in improving signal quality, decoding accuracy, and user adaptation. While challenges remain in long-term stability and user training, advances in adaptive algorithms and feedback mechanisms demonstrate the technology's growing viability for clinical applications. Recent innovations in electrode technology, AI architectures, and closed-loop systems, combined with emerging standardization frameworks, suggest accelerating progress toward widespread therapeutic use and human augmentation applications.

PMID:39719191 | DOI:10.1016/j.brainres.2024.149423

J Bone Joint Surg Am. 2024 Dec 24. doi: 10.2106/JBJS.24.00316. Online ahead of print.

ABSTRACT

BACKGROUND: Studies investigating constitutional alignment across various grades of osteoarthritis (OA) are limited. This study explored the distribution of Coronal Plane Alignment of the Knee (CPAK) types and associated radiographic parameters with increasing OA severity.

METHODS: In this retrospective cross-sectional study, 17,365 knees were analyzed using deep learning software for radiographic measurements. Knees were categorized on the basis of the Kellgren and Lawrence (KL) grade and CPAK type. Radiographic measurements were the hip-knee-ankle angle (HKAA), lateral distal femoral angle (LDFA), medial proximal tibial angle (MPTA), arithmetic HKAA (aHKA), joint line obliquity (JLO), and joint line convergence angle (JLCA). Age-stratified analysis was performed to differentiate the impact of age on OA severity.

RESULTS: A shift in the most common CPAK type from II to I was found with increasing KL grade (p < 0.05). Furthermore, there was a corresponding increase in LDFA and JLCA with increasing KL grade, while HKAA, MPTA, and aHKA decreased after KL grade 2. Age exhibited limited association with LDFA and MPTA, suggesting that OA severity is the dominant factor related to the CPAK distribution.

CONCLUSIONS: The study found a shift in CPAK type with worsening OA. It is possible that constitutional varus types are more susceptible to OA, or that their increased OA prevalence is related to anatomical changes. This analysis offers new insights into alterations in CPAK type that occur with OA and underscores the importance of understanding pre-arthritic anatomy when performing joint reconstruction.

LEVEL OF EVIDENCE: Prognostic Level III. See Instructions for Authors for a complete description of levels of evidence.

PMID:39719004 | DOI:10.2106/JBJS.24.00316

Opt Lett. 2025 Jan 1;50(1):189-192. doi: 10.1364/OL.540475.

ABSTRACT

Polarization and wavelength multiplexed metalenses address the bulkiness of traditional imaging systems. However, despite progress with numerical simulations and parameter scanning, the engineering complexity of classical methods highlights the urgent need for efficient deep learning approaches. This paper introduces a deep learning-driven inverse design model for polarization-multiplexed metalenses, employing propagation phase theory alongside spectral transfer learning to address chromatic dispersion challenges. The model facilitates the rapid design of metalenses with off-axis and dual-focus capabilities within a single wavelength. Numerical simulations reveal a focal length deviation of less than 5% and an average focusing efficiency of 43.3%. The integration of spectral transfer learning streamlines the design process, enabling multifunctional metalenses with enhanced full-color imaging and displacement measurement, thus advancing the field of metasurfaces.

PMID:39718885 | DOI:10.1364/OL.540475

Opt Lett. 2025 Jan 1;50(1):45-48. doi: 10.1364/OL.544716.

ABSTRACT

To overcome the limitations of optical coherence tomography (OCT) in imaging large-scale freeform objects, we propose a methodological framework that utilizes OCT as both a shape sensor and a tomographic imager in robotic scanning. Our approach integrates a deep-learning-based surface detection algorithm to counter OCT artifacts and an adaptive robotic arm pose adjustment algorithm for sensing and imaging uneven objects. We demonstrate the effectiveness and superiority of our method on various objects, achieving high-resolution, large-scale tomographic imaging that adeptly manages OCT artifacts and surface irregularities. We think this work may contribute to expanding the applicability of OCT in both medical and industrial scenarios.

PMID:39718851 | DOI:10.1364/OL.544716

J Med Internet Res. 2024 Dec 24;26:e53863. doi: 10.2196/53863.

ABSTRACT

BACKGROUND: Wearable technologies have become increasingly prominent in health care. However, intricate machine learning and deep learning algorithms often lead to the development of "black box" models, which lack transparency and comprehensibility for medical professionals and end users. In this context, the integration of explainable artificial intelligence (XAI) has emerged as a crucial solution. By providing insights into the inner workings of complex algorithms, XAI aims to foster trust and empower stakeholders to use wearable technologies responsibly.

OBJECTIVE: This paper aims to review the recent literature and explore the application of explainability in wearables. By examining how XAI can enhance the interpretability of generated data and models, this review sought to shed light on the possibilities that arise at the intersection of wearable technologies and XAI.

METHODS: We collected publications from ACM Digital Library, IEEE Xplore, PubMed, SpringerLink, JMIR, Nature, and Scopus. The eligible studies included technology-based research involving wearable devices, sensors, or mobile phones focused on explainability, machine learning, or deep learning and that used quantified self data in medical contexts. Only peer-reviewed articles, proceedings, or book chapters published in English between 2018 and 2022 were considered. We excluded duplicates, reviews, books, workshops, courses, tutorials, and talks. We analyzed 25 research papers to gain insights into the current state of explainability in wearables in the health care context.

RESULTS: Our findings revealed that wrist-worn wearables such as Fitbit and Empatica E4 are prevalent in health care applications. However, more emphasis must be placed on making the data generated by these devices explainable. Among various explainability methods, post hoc approaches stand out, with Shapley Additive Explanations as a prominent choice due to its adaptability. The outputs of explainability methods are commonly presented visually, often in the form of graphs or user-friendly reports. Nevertheless, our review highlights a limitation in user evaluation and underscores the importance of involving users in the development process.

CONCLUSIONS: The integration of XAI into wearable health care technologies is crucial to address the issue of black box models. While wrist-worn wearables are widespread, there is a notable gap in making the data they generate explainable. Post hoc methods such as Shapley Additive Explanations have gained traction for their adaptability in explaining complex algorithms visually. However, user evaluation remains an area in which improvement is needed, and involving users in the development process can contribute to more transparent and reliable artificial intelligence models in health care applications. Further research in this area is essential to enhance the transparency and trustworthiness of artificial intelligence models used in wearable health care technology.

PMID:39718820 | DOI:10.2196/53863

FASEB J. 2024 Dec 13;38(24):e70276. doi: 10.1096/fj.202401972RRR.

ABSTRACT

Multi-target strategy can serve as a valid treatment for severe acute respiratory syndrome coronavirus 2 (SARS-CoV-2), but existing drugs most focus on a single target. Thus, multi-target drugs that bind multiple sites simultaneously need to be urgently studied. Apigenin has antiviral and anti-inflammatory properties. Here, we comprehensively explored the potential effect and mechanism of apigenin in SARS-CoV-2 treatment by a network algorithm, deep learning, molecular docking, molecular dynamics (MD) simulation, and normal mode analysis (NMA). KATZ-based VDA prediction method (VDA-KATZ) indicated that apigenin may provide a latent drug therapy for SARS-CoV-2. Prediction of DTA using convolution model with self-attention (CSatDTA) showed potential binding affinity of apigenin with multiple targets of virus entry, assembly, and cytokine storms including cathepsin L (CTSL), membrane (M), envelope (E), Toll-like receptor 4 (TLR4), nuclear factor-kappa B (NF-κB), NOD-like receptor pyrin domain-containing protein 3 (NLRP3), apoptosis-associated speck-like protein (ASC), and cysteinyl aspartate-specific proteinase-1 (Caspase-1). Molecular docking indicated that apigenin could effectively bind these targets, and its stability was confirmed using MD simulation and NMA. Overall, apigenin is a multi-target inhibitor for the entry, assembly, and cytokine storms of SARS-CoV-2.

PMID:39718442 | DOI:10.1096/fj.202401972RRR

J Acoust Soc Am. 2024 Dec 1;156(6):4229-4241. doi: 10.1121/10.0034707.

ABSTRACT

This article presents a spatial environmental inversion scheme using broadband impulse signals with deep learning (DL) to model a single spatially-varying sediment layer over a fixed basement. The method is applied to data from the Seabed Characterization Experiment 2022 (SBCEX22) in the New England Mud-Patch (NEMP). Signal Underwater Sound (SUS) explosive charges generated impulsive signals recorded by a distributed array of bottom-moored hydrophones. The inversion scheme is first validated on a range-dependent synthetic test set simulating SBCEX22 conditions, then applied to experimental data to predict the lateral spatial structure of sediment sound speed and its ratio with the interfacial water sound speed. Traditional geoacoustic inversion requires significant computational resources. Here, a neural network enables rapid single-signal inversion, allowing the processing of 1836 signals along 722 tracks. The method is applied to both synthetic and experimental data. Results from experimental data suggest an increase in both absolute compressional sound speed and sound speed ratio from southwest to northeast in the NEMP, consistent with published coring surveys and geoacoustic inversion results. This approach demonstrates the potential of DL for efficient spatial geoacoustic inversion in shallow water environments.

PMID:39718359 | DOI:10.1121/10.0034707

Int J Med Robot. 2024 Dec;20(6):e70017. doi: 10.1002/rcs.70017.

ABSTRACT

BACKGROUND: Percutaneous puncture procedures, guided by image-guided robotic-assisted intervention (IGRI) systems, are susceptible to disruptions in patients' respiratory rhythm due to factors such as pain and psychological distress.

METHODS: We developed an IGRI system with a coded structured light camera and a binocular camera. Our system incorporates dual-pathway deep learning networks, combining convolutional long short-term memory (ConvLSTM) and point long short-term memory (PointLSTM) modules for real-time respiratory signal monitoring.

RESULTS: Our in-house dataset experiments demonstrate the superior performance of the proposed network in accuracy, precision, recall and F1 compared to separate use of PointLSTM and ConvLSTM for respiratory pattern classification.

CONCLUSION: In our IGRI system, a respiratory signal monitoring module was constructed with a binocular camera and dual-pathway deep learning networks. The integrated respiratory monitoring module provides a basis for the application of respiratory gating technology to IGRI systems and enhances surgical safety by security mechanisms.

PMID:39718347 | DOI:10.1002/rcs.70017

Phys Chem Chem Phys. 2024 Dec 24. doi: 10.1039/d4cp03238a. Online ahead of print.

ABSTRACT

Carboncones and fullerenes are exemplary π-conjugated carbon nanomaterials with unsaturated, positively curved surfaces, enabling the attachment of atoms or functional groups to enhance their physicochemical properties. However, predicting and understanding the addition patterns in functionalized carboncones and fullerenes are extremely challenging due to the formidable complexity of the regioselectivity exhibited in the adducts. Existing predictive models fall short in systems where the carbon molecular framework undergoes severe distortion upon high degrees of addition. Here, we propose an incremental deep learning approach to predict regioselectivity in the hydrogenation of carboncones and chlorination of fullerenes. Utilizing exclusively graph-based features, our deep neural network (DNN) models rely solely on atomic connectivity, without requiring 3D molecular coordinates as input or their iterative optimization. This advantage inherently avoids the risk of obtaining chemically unreasonable optimized structures, enabling the handling of highly distorted adducts. The DNN models allow us to study regioselectivity in hydrogenated carboncones of C70H20 and C62H16, accommodating up to at least 40 and 30 additional H atoms, respectively. Our approach also correctly predicts experimental addition patterns in C50Cl10 and C76Cln (n = 18, 24, and 28), whereas in the latter cases all other known methods have been proven unsuccessful. Compared to our previously developed topology-based models, the DNN's superior predictive power and generalization ability make it a promising tool for investigating complex addition patterns in similar chemical systems.

PMID:39718318 | DOI:10.1039/d4cp03238a

Med Phys. 2024 Dec 24. doi: 10.1002/mp.17601. Online ahead of print.

ABSTRACT

BACKGROUND: Modern radiation therapy techniques, such as intensity-modulated radiation therapy (IMRT) and volumetric-modulated arc therapy (VMAT), use complex fluence modulation strategies to achieve optimal patient dose distribution. Ensuring their accuracy necessitates rigorous patient-specific quality assurance (PSQA), traditionally done through pretreatment measurements with detector arrays. While effective, these methods are labor-intensive and time-consuming. Independent calculation-based methods leveraging advanced dose algorithms provide a reduced workload but cannot account for machine performance during delivery.

PURPOSE: This study introduces a novel unified deep-learning (DL) framework to enhance PSQA. The framework can combine the strengths of measurement- and calculation-based approaches.

METHODS: A comprehensive artificial training dataset, comprising 400,000 samples, was generated based on a rigorous mathematical model that describes the physical processes of radiation transport and interaction within both the medium and detector. This artificial data was used to pretrain the DL models, which were subsequently fine-tuned with a measured dataset of 400 IMRT segments to capture the machine-specific characteristics. Additional measurements of five IMRT plans were used as the unseen test dataset. Within the unified framework, a forward prediction model uses plan parameters to predict the measured dose distributions, while the backward prediction model reconstructs these parameters from actual measurements. The former enables a detailed control point (CP)-wise analysis. At the same time, the latter facilitates the reconstruction of treatment plans from the measurements and, subsequently, dose recalculation in the treatment planning system (TPS), as well as an independent second check software (VERIQA). This method has been tested with an OD 1600 SRS and an OD 1500 detector array with distinct spatial resolution and detector arrangement in combination with a dedicated upsampling model for the latter.

RESULTS: The final models could deliver highly accurate predictions of the measurements in the forward direction and the actual delivered plan parameters in the backward direction. In the forward direction, the test plans reached median gamma passing rates better than 94% for the OD 1600 SRS measurements. The upsampled OD 1500 measurements show similar performance with similar median gamma passing rates but a slightly higher variability. The 3D gamma passing rates from the comparisons between the original and reconstructed dose distributions in patients lie between 95.4% and 98.2% for the OD 1600 SRS and 94.7% and 98.5% for the interpolated OD 1500 measurements. The dose volume histograms (DVH) of the original and the reconstructed plans, recalculated in both the TPS and VERIQA, were evaluated for the organs at risk and targets based on clinical protocols and showed no clinically relevant deviations.

CONCLUSIONS: The flexibility of the implemented model architecture allows its adaptability to other delivery techniques and measurement modalities. Its utilization also reduces the requirements of the measurement devices. The proposed unified framework could play a decisive role in automating QA workflow, especially in the context of real-time adaptive radiation therapy (ART).

PMID:39718209 | DOI:10.1002/mp.17601

J Hum Lact. 2024 Dec 24:8903344241303867. doi: 10.1177/08903344241303867. Online ahead of print.

ABSTRACT

BACKGROUND: No research has been conducted on the use of deep learning for breastfeeding support.

RESEARCH AIM: This study aims to develop a nipple trauma evaluation system using deep learning.

METHODS: We used an exploratory data analysis approach to develop a deep-learning model for medical imaging. Employing object detection and classification, this Japanese study retrieved 753 images from a previous study. The classification protocol, based on the "seven signs of nipple trauma associated with breastfeeding," categorized the images into eight classes. For practical purposes, the eight original classes were consolidated into four broader categories: "None," "Minor," "Moderate," and "Severe," using data augmentation procedures that were consistent with the original classification system. The Precision, Recall, Overall Accuracy, and Area Under the Curve (AUC) were calculated, and the model's efficiency was evaluated using Frames Per Second (FPS).

RESULTS: The object detector's high mean average precision and frames per second rate for nipple and areola detection, confirmed exceptional accuracy. The eight-class image classifier returned notable AUC values, with fissures, peeling, purpura, and scabbing exceeding 0.8. The highest average recall and precision was for scabbing, and the lowest for blistering. The four-class classifier accurately predicted severe conditions, with an average AUC > 0.7, whereas categories without classifications and those deemed minor had lower recall and precision rates.

CONCLUSIONS: A sophisticated deep learning system detects and classifies nipple trauma automatically, potentially aiding breastfeeding caregivers through objective image assessment and operational improvements.

ABSTRACT IN JAPANESE: ： におけるのにするはわれていない。： は、をいたシステムのをとした。： では、をいたモデルをするため、データアプローチをいた。およびのをい、でわれたでされた753のをした。「にうの７」にづき、を8クラスにした。をし、4つのカテゴリ「なし」、「」、「」、「」の4つのカテゴリにし、のシステムにするデータをった。、、Overall Accuracy、AUC（）をし、モデルのはFPS（Frames Per Second）でした。： におけるいmAP（）とFPSがされ、およびのがされた。8クラスのは、、、、で0.8をえるなAUCがられた。とがもかったのはであり、でもかった。4クラスのはのをにし、AUCは0.7をえたが、なしやとされるカテゴリはとがいとなった。： をしたこのなシステムは、のとをでうことができ、なをじて、のとをサポートするなツールとなりる。Back Translation Completed by Hiroko Hongo, MSW, PhD, IBCLC.

PMID:39718190 | DOI:10.1177/08903344241303867

Front Plant Sci. 2024 Dec 2;15:1491706. doi: 10.3389/fpls.2024.1491706. eCollection 2024.

ABSTRACT

INTRODUCTION: Detecting strawberry growth stages is crucial for optimizing production management. Precise monitoring enables farmers to adjust management strategies based on the specific growth needs of strawberries, thereby improving yield and quality. However, dense planting patterns and complex environments within greenhouses present challenges for accurately detecting growth stages. Traditional methods that rely on large-scale equipment are impractical in confined spaces. Thus, the development of lightweight detection technologies suitable for portable devices has become essential.

METHODS: This paper presents SGSNet, a lightweight deep learning model designed for the fast and accurate detection of various strawberry growth stages. A comprehensive dataset covering the entire strawberry growth cycle is constructed to serve as the foundation for model training and testing. An innovative lightweight convolutional neural network, named GrowthNet, is designed as the backbone of SGSNet, facilitating efficient feature extraction while significantly reducing model parameters and computational complexity. The DySample adaptive upsampling structure is employed to dynamically adjust sampling point locations, thereby enhancing the detection capability for objects at different scales. The RepNCSPELAN4 module is optimized with the iRMB lightweight attention mechanism to achieve efficient multi-scale feature fusion, significantly improving the accuracy of detecting small targets from long-distance images. Finally, the Inner-IoU optimization loss function is applied to accelerate model convergence and enhance detection accuracy.

RESULTS: Testing results indicate that SGSNet performs exceptionally well across key metrics, achieving 98.83% precision, 99.45% recall, 99.14% F1 score, 99.50% mAP@0.5, and a loss value of 0.3534. It surpasses popular models such as Faster R-CNN, YOLOv10, and RT-DETR. Furthermore, SGSNet has a computational cost of only 14.7 GFLOPs and a parameter count as low as 5.86 million, demonstrating an effective balance between high performance and resource efficiency.

DISCUSSION: Lightweight deep learning model SGSNet not only exceeds the mainstream model in detection accuracy, but also greatly reduces the need for computing resources and is suitable for portable devices. In the future, the model can be extended to detect the growth stage of other crops, further advancing smart agricultural management.

PMID:39717733 | PMC:PMC11664550 | DOI:10.3389/fpls.2024.1491706

Front Plant Sci. 2024 Dec 9;15:1434163. doi: 10.3389/fpls.2024.1434163. eCollection 2024.

ABSTRACT

In order to achieve precise discrimination of leaf diseases in the Maize/Soybean intercropping system, i.e. leaf spot disease, rust disease, mixed leaf diseases, this study utilized hyperspectral imaging and deep learning algorithms for the classification of diseased leaves of maize and soybean. In the experiments, hyperspectral imaging equipment was used to collect hyperspectral images of leaves, and the regions of interest were extracted within the spectral range of 400 to 1000 nm. These regions included one or more infected areas on the leaves to obtain hyperspectral data. This approach aimed to enhance the accurate discrimination of different types of diseases, providing more effective technical support for the detection and control of crop diseases. The preprocessing of hyperspectral data involved four methods: Savitzky-Golay (SG), Standard Normal Variate (SNV), Multiplicative Scatter Correction (MSC) and 1st Derivative (1st Der). The 1st Der was found to be the optimal preprocessing method for hyperspectral data of maize and soybean diseases. Competitive Adaptive Reweighted Sampling (CARS), Successive Projections Algorithm (SPA) and Principal Component Analysis (PCA) were employed for feature extraction on the optimal preprocessed data. The Support Vector Machines (SVM), Bidirectional Long Short-Term Memory Network (BiLSTM) and Dung Beetle Optimization-Bidirectional Long Short-Term Memory Network (DBO-BiLSTM) were established for the discrimination of maize and soybean diseases. Comparative analysis indicated that, in the classification of maize and soybean diseases, the DBO-BiLSTM model based on the CARS extraction method (1st Der-CARS-DBO-BiLSTM) demonstrated the highest classification rate, reaching 98.7% on the test set. The research findings suggest that integrating hyperspectral imaging with both traditional and deep learning methods is a viable and effective approach for classifying diseases in the intercropping model of maize and soybean. These results offer a novel method and a theoretical foundation for the non-invasive, precise, and efficient identification of diseases in the intercropping model of maize and soybean, carrying positive implications for agricultural production.

PMID:39717723 | PMC:PMC11663666 | DOI:10.3389/fpls.2024.1434163