J Opt Soc Am A Opt Image Sci Vis. 2024 Sep 1;41(9):1702-1712. doi: 10.1364/JOSAA.529243.

ABSTRACT

Phase retrieval is about reconstructing original vectors/images from their Fourier intensity measurements. Deep learning methods have been introduced to solve the phase retrieval problem; however, most of the proposed approaches cannot improve the reconstruction quality of phase and amplitude of original images simultaneously. In this paper, we present a distributed amplitude and phase conditional generative adversarial network (D-APUCGAN) to achieve the high quality of phase and amplitude images at the same time. D-APUCGAN includes UCGAN, AUCGAN/PUCGAN, and APUCGAN. In this paper, we introduce the content loss function to constrain the similarity between the reconstructed image and the source image through the Frobenius norm and the total variation modulus. The proposed method promotes the quality of phase images better than just using amplitude images to train. The numerical experimental results show that the proposed cascade strategies are significantly effective and remarkable for natural and unnatural images, DIV2K testing datasets, MNIST dataset, and realistic data. Comparing with the conventional neural network methods, the evaluation metrics of PSNR and SSIM values in the proposed method are refined by about 2.25 dB and 0.18 at least, respectively.

PMID:39889034 | DOI:10.1364/JOSAA.529243

J Opt Soc Am A Opt Image Sci Vis. 2024 Oct 1;41(10):1987-1993. doi: 10.1364/JOSAA.531198.

ABSTRACT

The $\pm 1$st order diffraction of gratings is widely used in spectral analysis. However, when the incident light is non-monochromatic, the higher-order diffractions generated by traditional diffraction gratings are always superimposed on the useful first-order diffraction, complicating subsequent spectral decoding. In this paper, single-order diffraction gratings with a sinusoidal transmittance, called hexagonal diffraction gratings (HDGs), are designed using a convolutional neural network based on deep learning algorithm. The trained convolutional neural network can accurately retrieve the structural parameters of the HDGs. Simulation and experimental results confirm that the HDGs can effectively suppress higher-order diffractions above the third order. The intensity of third-order diffraction is reduced from 20% of the first-order diffraction to less than that of the background. This higher-order diffraction suppression property of the HDGs is promising for applications in fields such as synchrotron radiation, astrophysics, and soft x-ray lasers.

PMID:39889023 | DOI:10.1364/JOSAA.531198

J Opt Soc Am A Opt Image Sci Vis. 2024 Oct 1;41(10):1979-1986. doi: 10.1364/JOSAA.525577.

ABSTRACT

The invention of microscopy- and nanoscopy-based imaging technology opened up different research directions in life science. However, these technologies create the need for larger storage space, which has negative impacts on the environment. This scenario creates the need for storing such images in a memory-efficient way. Compact image representation (CIR) can solve the issue as it targets storing images in a memory-efficient way. Thus, in this work, we have designed a deep-learning-based CIR technique that selects key pixels using the guided U-Net (GU-Net) architecture [Asian Conference on Pattern Recognition, p. 317 (2023)], and then near-original images are constructed using a conditional generative adversarial network (GAN)-based architecture. The technique was evaluated on two microscopy- and two scanner-captured-image datasets and obtained good performance in terms of storage requirements and quality of the reconstructed images.

PMID:39889022 | DOI:10.1364/JOSAA.525577

Sci Adv. 2025 Jan 31;11(5):eads4224. doi: 10.1126/sciadv.ads4224. Epub 2025 Jan 31.

ABSTRACT

Optical neural networks (ONNs) are a promising computational alternative for deep learning due to their inherent massive parallelism for linear operations. However, the development of energy-efficient and highly parallel optical nonlinearities, a critical component in ONNs, remains an outstanding challenge. Here, we introduce a nonlinear optical microdevice array (NOMA) compatible with incoherent illumination by integrating the liquid crystal cell with silicon photodiodes at the single-pixel level. We fabricate NOMA with more than half a million pixels, each functioning as an optical analog of the rectified linear unit at ultralow switching energy down to 100 femtojoules per pixel. With NOMA, we demonstrate an optical multilayer neural network. Our work holds promise for large-scale and low-power deep ONNs, computer vision, and real-time optical image processing.

PMID:39888986 | DOI:10.1126/sciadv.ads4224

PLoS One. 2025 Jan 31;20(1):e0317990. doi: 10.1371/journal.pone.0317990. eCollection 2025.

ABSTRACT

The continuous development of the field of artificial intelligence, not only makes people's lives more convenient but also plays a role in the supervision and protection of people's lives and property safety. News of the fire is not uncommon, and fire has become the biggest hidden danger threatening the safety of public life and property. In this paper, a deep learning-based flame recognition method for complex scenes, GGSYOLOv5, is proposed. Firstly, a global attention mechanism (GAM) was added to the CSP1 module in the backbone part of the YOLOv5 network, and then a parameterless attention mechanism was added to the feature fusion part. Finally, packet random convolution (GSConv) was used to replace the original convolution at the output end. A large number of experiments show that the detection accuracy rate is 4.46% higher than the original algorithm, and the FPS is as high as 64.3, which can meet the real-time requirements. Moreover, the algorithm is deployed in the Jetson Nano embedded development board to build the flame detection system.

PMID:39888970 | DOI:10.1371/journal.pone.0317990

PLoS One. 2025 Jan 31;20(1):e0317996. doi: 10.1371/journal.pone.0317996. eCollection 2025.

ABSTRACT

Compared with histological examination of lung cancer, cytology is less invasive and provides better preservation of complete morphology and detail. However, traditional cytological diagnosis requires an experienced pathologist to evaluate all sections individually under a microscope, which is a time-consuming process with low interobserver consistency. With the development of deep neural networks, the You Only Look Once (YOLO) object-detection model has been recognized for its impressive speed and accuracy. Thus, in this study, we developed a model for intraoperative cytological segmentation of pulmonary lesions based on the YOLOv8 algorithm, which labels each instance by segmenting the image at the pixel level. The model achieved a mean pixel accuracy and mean intersection over union of 0.80 and 0.70, respectively, on the test set. At the image level, the accuracy and area under the receiver operating characteristic curve values for malignant and benign (or normal) lesions were 91.0% and 0.90, respectively. In addition, the model was deemed suitable for diagnosing pleural fluid cytology and bronchoalveolar lavage fluid cytology images. The model predictions were strongly correlated with pathologist diagnoses and the gold standard, indicating the model's ability to make clinical-level decisions during initial diagnosis. Thus, the proposed method is useful for rapidly localizing lung cancer cells based on microscopic images and outputting image interpretation results.

PMID:39888907 | DOI:10.1371/journal.pone.0317996

Med Biol Eng Comput. 2025 Jan 31. doi: 10.1007/s11517-025-03284-3. Online ahead of print.

ABSTRACT

The lumen centerline of the coronary artery allows vessel reconstruction used to detect stenoses and plaques. Discrete-action-based centerline extraction methods suffer from artifacts and plaques. This study aimed to develop a continuous-action-based method which performs more effectively in cases involving artifacts or plaques. A continuous-action deep reinforcement learning-based model was trained to predict the artery's direction and radius value. The model is based on an Actor-Critic architecture. The Actor learns a deterministic policy to output the actions made by an agent. These actions indicate the centerline's direction and radius value consecutively. The Critic learns a value function to evaluate the quality of the agent's actions. A novel DDR reward was introduced to measure the agent's action (both centerline extraction and radius estimate) at each step. The method achieved an average OV of 95.7%, OF of 93.6%, OT of 97.3%, and AI of 0.22 mm in 80 test data. In 53 cases with artifacts or plaques, it achieved an average OV of 95.0%, OF of 91.5%, OT of 96.7%, and AI of 0.23 mm. The 95% limits of agreement between the reference and estimated radius values were - 0.46 mm and 0.43 mm in the 80 test data. Experiments demonstrate that the Actor-Critic architecture can achieve efficient centerline extraction and radius estimate. Compared with discrete-action-based methods, our method performs more effectively in cases involving artifacts or plaques. The extracted centerlines and radius values allow accurate coronary artery reconstruction that facilitates the detection of stenoses and plaques.

PMID:39888471 | DOI:10.1007/s11517-025-03284-3

Med Biol Eng Comput. 2025 Jan 31. doi: 10.1007/s11517-025-03308-y. Online ahead of print.

ABSTRACT

Smart homes have the potential to enable remote monitoring of the health and well-being of older adults, leading to improved health outcomes and increased independence. However, current approaches only consider a limited set of daily activities and do not combine data from individuals. In this work, we propose the use of deep learning techniques to model behavior at the population level and detect significant deviations (i.e., anomalies) while taking into account the whole set of daily activities (41). We detect and visualize daily routine patterns, train a set of recurrent neural networks for behavior modelling with next-day prediction, and model errors with a normal distribution to identify significant deviations while considering the temporal component. Clustering of daily routines achieves a silhouette score of 0.18 and the best model obtains a mean squared error in next day routine prediction of 4.38%. The mean number of deviated activities for the anomalies in the train and test set are 3.6 and 3.0, respectively, with more than 60% of anomalies involving three or more deviated activities in the test set. The methodology is scalable and can incorporate additional activities into the analysis.

PMID:39888470 | DOI:10.1007/s11517-025-03308-y

Clin Oral Investig. 2025 Jan 31;29(2):101. doi: 10.1007/s00784-025-06156-0.

ABSTRACT

OBJECTIVE: This study aimed to apply the DeepLabv3 + model and compare it with the U-Net model in terms of detecting and segmenting apical lesions on panoramic radiography.

METHODS: 260 panoramic images that contain apical lesions in different regions were collected and randomly divided into training and test datasets. All images were manually annotated for apical lesions using Computer Vision Annotation Tool software by two independent dental radiologists and a master reviewer. The DeepLabv3 + model, one of the state-of-the-art deep semantic segmentation models, was utilized using Python programming language and the TensorFlow library and applied to the prepared datasets. The model was compared with the U-Net model applied to apical lesions and other medical image segmentation problems in the literature.

RESULTS: The DeepLabv3 + and U-Net models were applied to the same datasets with the same hyper-parameters. The AUC and recall results of the DeepLabv3 + were 29.96% and 61.06% better than the U-Net model. However, the U-Net model gets 69.17% and 25.55% better precision and F1-score results than the DeepLabv3 + model. The difference in the IoU results of the models was not statistically significant.

CONCLUSIONS: This paper comprehensively evaluated the DeepLabv3 + model and compared it with the U-Net model. Our experimental findings indicated that DeepLabv3 + outperforms the U-Net model by a substantial margin for both AUC and recall metrics. According to those results, for detecting apical lesions, we encourage researchers to use and improve the DeepLabv3 + model.

CLINICAL RELEVANCE: The DeepLabv3 + model has the poten tial to improve clinical diagnosis and treatment planning and save time in the clinic.

PMID:39888441 | DOI:10.1007/s00784-025-06156-0

Microsc Res Tech. 2025 Jan 31. doi: 10.1002/jemt.24792. Online ahead of print.

ABSTRACT

Skin cancer is recognized as one of the most harmful cancers worldwide. Early detection of this cancer is an effective measure for treating the disease efficiently. Traditional skin cancer detection methods face scalability challenges and overfitting issues. To address these complexities, this study proposes a random cat swarm optimization (CSO)with an ensemble convolutional neural network (RCS-ECNN) method to categorize the different stages of skin cancer. In this study, two deep learning classifiers, deep neural network (DNN) and Keras DNN (KDNN), are utilized to identify the stages of skin cancer. In this method, an effective preprocessing phase is presented to simplify the classification process. The optimal features are selected using the feature extraction phase. Then, the GrabCut algorithm is employed to carry out the segmentation process. Also, the CSO is employed to enhance the effectiveness of the method. The HAM10000 and ISIC datasets are utilized to evaluate the RCS-ECNN method. The RCS-ECNN method achieved an accuracy of 99.56%, a recall of 99.66%, a specificity value of 99.254%, a precision value of 99.18%, and an F1-score value of 98.545%, respectively. The experimental results demonstrated that the RCS-ECNN method outperforms the existing techniques.

PMID:39888306 | DOI:10.1002/jemt.24792

Proteomics. 2025 Jan 31:e202400321. doi: 10.1002/pmic.202400321. Online ahead of print.

ABSTRACT

Metaproteomics enables the large-scale characterization of microbial community proteins, offering crucial insights into their taxonomic composition, functional activities, and interactions within their environments. By directly analyzing proteins, metaproteomics offers insights into community phenotypes and the roles individual members play in diverse ecosystems. Although database-dependent search engines are commonly used for peptide identification, they rely on pre-existing protein databases, which can be limiting for complex, poorly characterized microbiomes. De novo sequencing presents a promising alternative, which derives peptide sequences directly from mass spectra without requiring a database. Over time, this approach has evolved from manual annotation to advanced graph-based, tag-based, and deep learning-based methods, significantly improving the accuracy of peptide identification. This Viewpoint explores the evolution, advantages, limitations, and future opportunities of de novo sequencing in metaproteomics. We highlight recent technological advancements that have improved its potential for detecting unsequenced species and for providing deeper functional insights into microbial communities.

PMID:39888246 | DOI:10.1002/pmic.202400321

Adv Sci (Weinh). 2025 Jan 30:e2413571. doi: 10.1002/advs.202413571. Online ahead of print.

ABSTRACT

DNA methylation plays a critical role in gene regulation, affecting cellular differentiation and disease progression, particularly in non-coding regions. However, predicting the epigenetic consequences of non-coding mutations at single-cell resolution remains a challenge. Existing tools have limited prediction capacity and struggle to capture dynamic, cell-type-specific regulatory changes that are crucial for understanding disease mechanisms. Here, Methven, a deep learning framework designed is presented to predict the effects of non-coding mutations on DNA methylation at single-cell resolution. Methven integrates DNA sequence with single-cell ATAC-seq data and models SNP-CpG interactions over 100 kbp genomic distances. By using a divide-and-conquer approach, Methven accurately predicts both short- and long-range regulatory interactions and leverages the pre-trained DNA language model for enhanced precision in classification and regression tasks. Methven outperforms existing methods and demonstrates robust generalizability to monocyte datasets. Importantly, it identifies CpG sites associated with rheumatoid arthritis, revealing key pathways involved in immune regulation and disease progression. Methven's ability to detect progressive epigenetic changes provides crucial insights into gene regulation in complex diseases. These findings demonstrate Methven's potential as a powerful tool for basic research and clinical applications, advancing this understanding of non-coding mutations and their role in disease, while offering new opportunities for personalized medicine.

PMID:39888214 | DOI:10.1002/advs.202413571

Oral Dis. 2025 Jan 30. doi: 10.1111/odi.15269. Online ahead of print.

ABSTRACT

OBJECTIVES: Utilizing a deep learning approach is an emerging trend to improve the efficiency of periodontitis diagnosis and classification. This study aimed to use an object detection model to automatically annotate the anatomic structure and subsequently classify the stages of radiographic bone loss (RBL).

MATERIALS AND METHODS: In all, 558 panoramic radiographs were cropped to 7359 pieces of individual teeth. The detection performance of the model was assessed using mean average precision (mAP), root mean squared error (RMSE). The classification performance was evaluated using accuracy, precision, recall, and F1 score. Additionally, receiver operating characteristic (ROC) curves and confusion matrices were presented, and the area under the ROC curve (AUC) was calculated.

RESULTS: The mAP was 0.88 when the difference between the ground truth and prediction was 10 pixels, and 0.99 when the difference was 25 pixels. For all images, the mean RMSE was 7.30 pixels. Overall, the accuracy, precision, recall, F1 score, and micro-average AUC of the prediction were 0.72, 0.76, 0.64, 0.68, and 0.79, respectively.

CONCLUSIONS: The current model is reliable in assisting with the detection and staging of radiographic bone levels.

PMID:39888112 | DOI:10.1111/odi.15269

Mol Inform. 2025 Feb;44(2):e202400039. doi: 10.1002/minf.202400039.

ABSTRACT

Currently, numerous metrics allow chemists and computational chemists to refine and filter libraries of virtual molecules in order to prioritize their synthesis. Some of the most commonly used metrics and models are QSAR models, docking scores, diverse druggability metrics, and synthetic feasibility scores to name only a few. To our knowledge, among the known metrics, a function which estimates the price of a novel virtual molecule and which takes into account the availability and price of starting materials has not been considered before in literature. Being able to make such a prediction could improve and accelerate the decision-making process related to the cost-of-goods. Taking advantage of recent advances in the field of Computer Aided Synthetic Planning (CASP), we decided to investigate if the predicted retrosynthetic pathways of a given molecule and the prices of its associated starting materials could be good features to predict the price of that compound. In this work, we present a deep learning model, RetroPriceNet, that predicts the price of molecules using their predicted synthetic pathways. On a holdout test set, the model achieves better performance than the state-of-the-art model. The developed approach takes into account the synthetic feasibility of molecules and the availability and prices of the starting materials.

PMID:39887833 | DOI:10.1002/minf.202400039

J Magn Reson Imaging. 2025 Jan 29. doi: 10.1002/jmri.29682. Online ahead of print.

ABSTRACT

BACKGROUND: Conventional quantitative MRI (qMRI) scan is time-consuming and highly sensitive to movements, posing great challenges for quantitative images of individuals with involuntary movements, such as Huntington's disease (HD).

PURPOSE: To evaluate the potential of our developed ultra-fast qMRI technique, multiple overlapping-echo detachment (MOLED), in overcoming involuntary head motion and its capacity to quantitatively assess tissue changes in HD.

STUDY TYPE: Prospective.

PHANTOM/SUBJECTS: A phantom comprising 13 tubes of MnCl2 at varying concentrations, 5 healthy volunteers (male/female: 1/4), 22 HD patients (male/female: 14/8) and 27 healthy controls (male/female: 15/12).

FIELD STRENGTH/SEQUENCE: 3.0 T. MOLED-T2 sequence, MOLED-T2* sequence, T2-weighted spin-echo sequence, T1-weighted gradient echo sequence, and T2-dark-fluid sequence.

ASSESSMENT: T1-weighted images were reconstructed into high-resolution images, followed by segmentation to delineate regions of interest (ROIs). Subsequently, the MOLED T2 and T2* maps were aligned with the high-resolution images, and the ROIs were transformed into the MOLED image space using the transformation matrix and warp field. Finally, T2 and T2* values were extracted from the MOLED relaxation maps.

STATISTICAL TESTS: Bland-Altman analysis, independent t test, Mann-Whitney U test, Pearson correlation analysis, and Spearman correlation analysis, P < 0.05 was considered statistically significant.

RESULTS: MOLED-T2 and MOLED-T2* sequences demonstrated good accuracy (Meandiff = - 0.20%, SDdiff = 1.05%, and Meandiff = -1.73%, SDdiff = 10.98%, respectively), and good repeatability (average intraclass correlation coefficient: 0.856 and 0.853, respectively). More important, MOLED T2 and T2* maps remained artifact-free across all HD patients, even in the presence of apparent head motions. Moreover, there were significant differences in T2 and T2* values across multiple ROIs between HD and controls.

DATA CONCLUSION: The ultra-fast scanning capabilities of MOLED effectively mitigate the impact of head movements, offering a robust solution for quantitative imaging in HD. Moreover, T2 and T2* values derived from MOLED provide powerful capabilities for quantifying tissue changes.

PLAIN LANGUAGE SUMMARY: Quantitative MRI scan is time-consuming and sensitive to movements. Consequently, obtaining quantitative images is challenging for patients with involuntary movements, such as those with Huntington's Disease (HD). In response, a newly developed MOLED technique has been introduced, promising to resist motion through ultra-fast scan. This technique has demonstrated excellent accuracy and reproducibility and importantly all HD patient's MOLED maps remained artifacts-free. Additionally, there were significant differences in T2 and T2∗ values across ROIs between HD and controls. The robust resistance of MOLED to motion makes it particularly suitable for quantitative assessments in patients prone to involuntary movements.

LEVEL OF EVIDENCE: 2 TECHNICAL EFFICACY: Stage 1.

PMID:39887812 | DOI:10.1002/jmri.29682

Med Phys. 2025 Jan 30. doi: 10.1002/mp.17651. Online ahead of print.

ABSTRACT

BACKGROUND: Deep learning image reconstruction (DLIR) algorithms allow strong noise reduction while preserving noise texture, which may potentially improve hypervascular focal liver lesions.

PURPOSE: To assess the impact of DLIR on image quality (IQ) and detectability of simulated hypervascular hepatocellular carcinoma (HCC) in fast kV-switching dual-energy CT (DECT).

METHODS: An anthropomorphic phantom of a standard patient morphology (body mass index of 23 kg m-2) with customized liver, including mimickers of hypervascular lesions in both late arterial phase (AP) and portal venous phase (PVP) enhancement, was scanned on a DECT. Virtual monoenergetic images were reconstructed from raw data at four energy levels (40/50/60/70 keV) using filtered back-projection (FBP), adaptive statistical iterative reconstruction-V 50% and 100% (ASIRV-50 and ASIRV-100), DLIR low (DLIR-L), medium (DLIR-M), and high (DLIR-H). The contrast between the lesion and the liver parenchyma, the noise magnitude, the average and peak frequencies (favg and fpeak) of the noise power spectrum (NPS) reflecting noise texture, and the task-based measure of the modulation transfer function (MTFtask) were measured to evaluate spatial resolution. A detectability index (d') was computed to model the detection of hypervascular lesions in both AP and PVP. Metrics were compared between reconstructions and between energy levels using a Friedman test with follow-up post-hoc multiple comparison.

RESULTS: Lesion-to-liver contrast significantly increased with decreasing energy level in both AP and PVP (p ≤ 0.042) but was not affected by reconstruction algorithm (p ≥ 0.57). Overall, noise magnitude increased with decreasing energy levels and was the lowest with ASIRV-100 at all energy levels in both AP and PVP (p ≤ 0.01) and significantly lower with DLIR-M and DLIR-H reconstructions compared to ASIRV-50 and DLIR-L (p < 0.001). For all reconstructions, noise texture within the liver tended to get smoother with decreasing energy; favg significantly shifted towards lower frequencies from 70 to 40 keV (p ≤ 0.01). Noise texture was the smoothest with ASIRV-100 (p < 0.001) while DLIR-L had the noise texture closer to the one of FBP. The spatial resolution was not significantly affected by the energy level, but it was degraded when increasing the level of ASIRV and DLIR. For all reconstructions, the detectability indices increased with decreasing energy levels and peaked at 40 and 50 keV in AP and PVP, respectively. In both AP and PVP, the highest d' values were observed with ASIRV-100 and DLIR-H, whatever the energy level studied (p ≤ 0.01) without statistical significance between those two reconstructions.

CONCLUSIONS: Compared to the routinely used level of iterative reconstruction, DLIR reduces noise without consequential noise texture modification, and may improve the detectability of hypervascular liver lesions while enabling the use of lower energy virtual monoenergetic images. The optimal energy level and DLIR level may depend on the lesion enhancement.

PMID:39887750 | DOI:10.1002/mp.17651

Front Plant Sci. 2025 Jan 16;15:1442686. doi: 10.3389/fpls.2024.1442686. eCollection 2024.

ABSTRACT

The stomatal phenotype is a crucial microscopic characteristic of the leaf surface, and modulating the stomata of maize leaves can enhance photosynthetic carbon assimilation and water use efficiency, thereby playing a vital role in maize yield formation. The evolving imaging and image processing technologies offer effective tools for precise analysis of stomatal phenotypes. This study employed Jingnongke 728 and its parental inbred to capture stomatal images from various leaf positions and abaxial surfaces during key reproductive stages using rapid scanning electron microscopy. We uesd a target detection and image segmentation approach based on YOLOv5s and Unet to efficiently obtain 11 phenotypic traits encompassing stomatal count, shape, and distribution. Manual validation revealed high detection accuracies for stomatal density, width, and length, with R2 values of 0.92, 0.97, and 0.95, respectively. Phenotypic analyses indicated a significant positive correlation between stomatal density and the percentage of guard cells and pore area (r=0.36), and a negative correlation with stomatal area and subsidiary cell area (r=-0.34 and -0.46). Additionally, stomatal traits exhibited notable variations with reproductive stages and leaf layers. Specifically, at the monocot scale, stomatal density increased from 74.35 to 87.19 Counts/mm2 from lower to upper leaf layers. Concurrently, the stomatal shape shifted from sub-circular (stomatal roundness = 0.64) to narrow and elongated (stomatal roundness = 0.63). Throughout the growth cycle, stomatal density remained stable during vegetative growth, decreased during reproductive growth with smaller size and narrower shape, and continued to decline while increasing in size and tending towards a rounded shape during senescence. Remarkably, hybrid 728 differed notably from its parents in stomatal phenotype, particularly during senescence. Moreover, the stomatal density of the hybrids showed negative super parental heterosis (heterosis rate = -0.09), whereas stomatal dimensions exhibited positive super parental heterosis, generally resembling the parent MC01. This investigation unveils the dynamic variations in maize stomatal phenotypes, bolstering genetic analyses and targeted improvements in maize, and presenting a novel technological instrument for plant phenotype studies.

PMID:39886688 | PMC:PMC11779725 | DOI:10.3389/fpls.2024.1442686

Front Plant Sci. 2025 Jan 16;15:1457360. doi: 10.3389/fpls.2024.1457360. eCollection 2024.

ABSTRACT

With the rapid advancement of plant phenotyping research, understanding plant genetic information and growth trends has become crucial. Measuring seedling length is a key criterion for assessing seed viability, but traditional ruler-based methods are time-consuming and labor-intensive. To address these limitations, we propose an efficient deep learning approach to enhance plant seedling phenotyping analysis. We improved the DeepLabv3+ model, naming it DFMA, and introduced a novel ASPP structure, PSPA-ASPP. On our self-constructed rice seedling dataset, the model achieved a mean Intersection over Union (mIoU) of 81.72%. On publicly available datasets, including Arabidopsis thaliana, Brachypodium distachyon, and Sinapis alba, detection scores reached 87.69%, 91.07%, and 66.44%, respectively, outperforming existing models. The model generates detailed segmentation masks, capturing structures such as the embryonic shoot, axis, and root, while a seedling length measurement algorithm provides precise parameters for component development. This approach offers a comprehensive, automated solution, improving phenotyping analysis efficiency and addressing the challenges of traditional methods.

PMID:39886686 | PMC:PMC11779734 | DOI:10.3389/fpls.2024.1457360

Front Plant Sci. 2025 Jan 16;15:1458978. doi: 10.3389/fpls.2024.1458978. eCollection 2024.

ABSTRACT

Hyperspectral images are rich in spectral and spatial information, providing a detailed and comprehensive description of objects, which makes hyperspectral image analysis technology essential in intelligent agriculture. With various corn seed varieties exhibiting significant internal structural differences, accurate classification is crucial for planting, monitoring, and consumption. However, due to the large volume and complex features of hyperspectral corn image data, existing methods often fall short in feature extraction and utilization, leading to low classification accuracy. To address these issues, this paper proposes a spectral-spatial attention transformer network (SSATNet) for hyperspectral corn image classification. Specifically, SSATNet utilizes 3D and 2D convolutions to effectively extract local spatial, spectral, and textural features from the data while incorporating spectral and spatial morphological structures to understand the internal structure of the data better. Additionally, a transformer encoder with cross-attention extracts and refines feature information from a global perspective. Finally, a classifier generates the prediction results. Compared to existing state-of-the-art classification methods, our model performs better on the hyperspectral corn image dataset, demonstrating its effectiveness.

PMID:39886680 | PMC:PMC11781253 | DOI:10.3389/fpls.2024.1458978

Proc IEEE Int Symp Biomed Imaging. 2024 May;2024. doi: 10.1109/ISBI56570.2024.10635138. Epub 2024 Aug 22.

ABSTRACT

Physics-driven deep learning (PD-DL) methods have gained popularity for improved reconstruction of fast MRI scans. Though supervised learning has been used in early works, there has been a recent interest in unsupervised learning methods for training PD-DL. In this work, we take inspiration from statistical image processing and compressed sensing (CS), and propose a novel convex loss function as an alternative learning strategy. Our loss function evaluates the compressibility of the output image while ensuring data fidelity to assess the quality of reconstruction in versatile settings, including supervised, unsupervised, and zero-shot scenarios. In particular, we leverage the reweighted l 1 norm that has been shown to approximate the l 0 norm for quality evaluation. Results show that the PD-DL networks trained with the proposed loss formulation outperform conventional methods, while maintaining similar quality to PD-DL models trained using existing supervised and unsupervised techniques.

PMID:39886655 | PMC:PMC11779509 | DOI:10.1109/ISBI56570.2024.10635138