Nat Methods. 2024 Nov 21. doi: 10.1038/s41592-024-02487-0. Online ahead of print.

ABSTRACT

Accurate prediction of RNA three-dimensional (3D) structures remains an unsolved challenge. Determining RNA 3D structures is crucial for understanding their functions and informing RNA-targeting drug development and synthetic biology design. The structural flexibility of RNA, which leads to the scarcity of experimentally determined data, complicates computational prediction efforts. Here we present RhoFold+, an RNA language model-based deep learning method that accurately predicts 3D structures of single-chain RNAs from sequences. By integrating an RNA language model pretrained on ~23.7 million RNA sequences and leveraging techniques to address data scarcity, RhoFold+ offers a fully automated end-to-end pipeline for RNA 3D structure prediction. Retrospective evaluations on RNA-Puzzles and CASP15 natural RNA targets demonstrate the superiority of RhoFold+ over existing methods, including human expert groups. Its efficacy and generalizability are further validated through cross-family and cross-type assessments, as well as time-censored benchmarks. Additionally, RhoFold+ predicts RNA secondary structures and interhelical angles, providing empirically verifiable features that broaden its applicability to RNA structure and function studies.

PMID:39572716 | DOI:10.1038/s41592-024-02487-0

Sci Rep. 2024 Nov 21;14(1):28904. doi: 10.1038/s41598-024-80199-3.

ABSTRACT

Deep learning-based computer vision systems have become powerful tools for automated and cost-effective pavement distress detection, essential for efficient road maintenance. Current methods focus primarily on developing supervised learning architectures, which are limited by the scarcity of annotated image datasets. The use of data augmentation with synthetic images created by generative models to improve these supervised systems is not widely explored. The few studies that do focus on generative architectures are mostly non-conditional, requiring extra labeling, and typically address only road crack defects while aiming to improve classification models rather than object detection. This study introduces AsphaltGAN, a novel class-conditional Generative Adversarial Network with attention mechanisms, designed to augment datasets with various rare road defects to enhance object detection. An in-depth analysis evaluates the impact of different loss functions and hyperparameter tuning. The optimized AsphaltGAN outperforms state-of-the-art generative architectures on public datasets. Additionally, a new workflow is proposed to improve object detection models using synthetic road images. The augmented datasets significantly improve the object detection metrics of You Only Look Once version 8 by 33.0%, 3.8%, 46.3%, and 51.8% on the Road Damage Detection 2022 dataset, Crack Dataset, Asphalt Pavement Detection Dataset, and Crack Surface Dataset, respectively.

PMID:39572659 | DOI:10.1038/s41598-024-80199-3

Sci Rep. 2024 Nov 21;14(1):28850. doi: 10.1038/s41598-024-80087-w.

ABSTRACT

Skin Cancer, which leads to a large number of deaths annually, has been extensively considered as the most lethal tumor around the world. Accurate detection of skin cancer in its early stage can significantly raise the survival rate of patients and reduce the burden on public health. Currently, the diagnosis of skin cancer relies heavily on human visual system for screening and dermoscopy. However, manual inspection is laborious, time-consuming, and error-prone. In consequence, the development of an automatic machine vision algorithm for skin cancer classification turns into imperative. Various machine learning techniques have been presented for the last few years. Although these methods have yielded promising outcome in skin cancer detection and recognition, there is still a certain gap in machine learning algorithms and clinical applications. To enhance the performance of classification, this study proposes a novel deep learning model for discriminating clinical skin cancer images. The proposed model incorporates a convolutional neural network for extracting local receptive field information and a novel attention mechanism for revealing the global associations within an image. Experimental results of the proposed approach demonstrate its superiority over the state-of-the-art algorithms on the publicly available dataset International Skin Imaging Collaboration 2019 (ISIC-2019) in terms of Precision, Recall, F1-score. From the experimental results, it can be observed that the proposed approach is a potentially valuable instrument for skin cancer classification.

PMID:39572649 | DOI:10.1038/s41598-024-80087-w

Sci Rep. 2024 Nov 21;14(1):28822. doi: 10.1038/s41598-024-79435-7.

ABSTRACT

The preparation of sintered NdFeB magnets is complex, time-consuming, and costly. Data-driven machine learning methods can enhance the efficiency of material synthesis and performance optimization. Traditional machine learning models based on mathematical and statistical principles are effective for structured data and offer high interpretability. However, as the scale and dimensionality of the data increase, the computational complexity of models rises dramatically, making hyperparameter tuning more challenging. By contrast, neural network models possess strong nonlinear modeling capabilities for handling large-scale data, but their decision-making and inferential processes remain opaque. To enhance interpretability of neural network, we collected 1,200 high-quality experimental data points and developed a multi-head attention regression model by integrating an attention mechanism into the neural network. The model enables parallel data processing, accelerates both training and inference speed, and reduces reliance on feature engineering and hyperparameter tuning. The coefficients of determination for remanence and coercivity are 0.97 and 0.84, respectively. This study offers new insights into machine learning-based modeling of structure-property relationships in materials and has potential to advance the research of multimodal NdFeB magnet models.

PMID:39572633 | DOI:10.1038/s41598-024-79435-7

Sci Rep. 2024 Nov 21;14(1):28905. doi: 10.1038/s41598-024-80371-9.

ABSTRACT

This study presents a data-driven battery emulator using long short-term memory deep learning models to predict the charge-discharge behaviour of lithium-ion batteries (LIBs). This study aimed to reduce the economic costs and time associated with the fabrication of large-scale automotive prototype batteries by emulating their performance using smaller laboratory-produced batteries. Two types of datasets were targeted: simulation data from the Dualfoil model and experimental data from liquid-based LIBs. These datasets were used to accurately predict the voltage profiles from the arbitrary inputs of various galvanostatic charge-discharge schedules. The results demonstrated high prediction accuracy, with the coefficient of determination scores reaching 0.98 and 0.97 for test datasets obtained from the simulation and experiments, respectively. The study also confirmed the significance of state-of-charge descriptors and inferred that a robust model performance could be achieved with as few as five charge-discharge training datasets. This study concludes that data-driven emulation using machine learning can significantly accelerate the battery development process, providing a powerful tool for reducing the time and economic costs associated with the production of large-scale prototype batteries.

PMID:39572616 | DOI:10.1038/s41598-024-80371-9

Sci Data. 2024 Nov 21;11(1):1265. doi: 10.1038/s41597-024-04114-1.

ABSTRACT

The rapid advancement of deep learning has enabled Brain-Computer Interfaces (BCIs) technology, particularly neural decoding techniques, to achieve higher accuracy and deeper levels of interpretation. Interest in decoding imagined speech has significantly increased because its concept akin to "mind reading". However, previous studies on decoding neural language have predominantly focused on brain activity patterns during human reading. The absence of imagined speech electroencephalography (EEG) datasets has constrained further research in this field. We present the Chinese Imagined Speech Corpus (Chisco), including over 20,000 sentences of high-density EEG recordings of imagined speech from healthy adults. Each subject's EEG data exceeds 900 minutes, representing the largest dataset per individual currently available for decoding neural language to date. Furthermore, the experimental stimuli include over 6,000 everyday phrases across 39 semantic categories, covering nearly all aspects of daily language. We believe that Chisco represents a valuable resource for the fields of BCIs, facilitating the development of more user-friendly BCIs.

PMID:39572577 | DOI:10.1038/s41597-024-04114-1

J Chem Inf Model. 2024 Nov 21. doi: 10.1021/acs.jcim.4c01139. Online ahead of print.

ABSTRACT

Drug discovery and development is a complex and costly process, with a substantial portion of the expense dedicated to characterizing the absorption, distribution, metabolism, excretion, and toxicity (ADMET) properties of new drug candidates. While the advent of deep learning and molecular graph neural networks (GNNs) has significantly enhanced in silico ADMET prediction capabilities, reliably quantifying prediction uncertainty remains a critical challenge. The performance of GNNs is influenced by both the volume and the quality of the data. Hence, determining the reliability and extent of a prediction is as crucial as achieving accurate predictions, especially for out-of-domain (OoD) compounds. This paper introduces a novel GNN model called conformalized fusion regression (CFR). CFR combined a GNN model with a joint mean-quantile regression loss and an ensemble-based conformal prediction (CP) method. Through rigorous evaluation across various ADMET tasks, we demonstrate that our framework provides accurate predictions, reliable probability calibration, and high-quality prediction intervals, outperforming existing uncertainty quantification methods.

PMID:39571080 | DOI:10.1021/acs.jcim.4c01139

PLoS One. 2024 Nov 21;19(11):e0311601. doi: 10.1371/journal.pone.0311601. eCollection 2024.

ABSTRACT

The integration of IoT (Internet of Things) devices has emerged as a technical cornerstone in the landscape of modern agriculture, revolutionising the way farming practises are viewed and managed. Smart farming, enabled by interconnected sensors and technologies, has surpassed traditional methods, giving farmers real-time, granular information into their farms. These Internet of Things devices are responsible for collecting and sending greenhouse data (temperature, humidity, and soil moisture) for the required destination, to provide a comprehensive awareness of environmental factors critical to crop growth. Therefore, ensuring that the received data are accurate is a challenge, thus this paper investigates the optimization of Agriculture IoT communication, proposing a complete strategy for improving data transmission efficiency within smart farming ecosystems. The proposed model intends to maximize energy efficiency and data throughput in the context of essential agricultural factors by using Lagrange optimization and a Deep Convolutional Neural Network (DCNN). The paper focus on the ideal communication required distance between IoT sensors that measure humidity, temperature, and water levels and central control systems. The investigation emphasizes the critical necessity of these data points in guaranteeing crop health and vitality. The proposed technique strives to improve the performance of agricultural IoT communication networks through the integration of mathematical optimization and cutting-edge deep learning. This paradigm change emphasizes the inherent link between precise achievable data rate and energy efficiency, resulting in resilient agricultural ecosystems capable of adjusting to dynamic environmental conditions for optimal crop output and health.

PMID:39570960 | DOI:10.1371/journal.pone.0311601

PLoS One. 2024 Nov 21;19(11):e0309706. doi: 10.1371/journal.pone.0309706. eCollection 2024.

ABSTRACT

Motor imagery (MI)-electroencephalography (EEG) decoding plays an important role in brain-computer interface (BCI), which enables motor-disabled patients to communicate with external world via manipulating smart equipment. Currently, deep learning (DL)-based methods are popular for EEG decoding. Whereas the utilization efficiency of EEG features in frequency and temporal domain is not sufficient, which results in poor MI classification performance. To address this issue, an EEG-based MI classification model based on a frequency enhancement module, a deformable convolutional network, and a crop module (FDCN-C) is proposed. Firstly, the frequency enhancement module is innovatively designed to address the issue of extracting frequency information. It utilizes convolution kernels at continuous time scales to extract features across different frequency bands. These features are screened by calculating attention and integrated into the original EEG data. Secondly, for temporal feature extraction, a deformable convolution network is employed to enhance feature extraction capabilities, utilizing offset parameters to modulate the convolution kernel size. In spatial domain, a one-dimensional convolution layer is designed to integrate all channel information. Finally, a dilated convolution is used to form a crop classification module, wherein the diverse receptive fields of the EEG data are computed multiple times. Two public datasets are employed to verify the proposed FDCN-C model, the classification accuracy obtained from the proposed model is greater than that of state-of-the-art methods. The model's accuracy has improved by 14.01% compared to the baseline model, and the ablation study has confirmed the effectiveness of each module in the model.

PMID:39570849 | DOI:10.1371/journal.pone.0309706

PLoS One. 2024 Nov 21;19(11):e0313261. doi: 10.1371/journal.pone.0313261. eCollection 2024.

ABSTRACT

Motor imagery (MI) classification has been commonly employed in making brain-computer interfaces (BCI) to manage the outside tools as a substitute neural muscular path. Effectual MI classification in BCI improves communication and mobility for people with a breakdown or motor damage, delivering a bridge between the brain's intentions and exterior actions. Employing electroencephalography (EEG) or aggressive neural recordings, machine learning (ML) methods are used to interpret patterns of brain action linked with motor image tasks. These models frequently depend upon models like support vector machine (SVM) or deep learning (DL) to distinguish among dissimilar MI classes, such as visualizing left or right limb actions. This procedure allows individuals, particularly those with motor disabilities, to utilize their opinions to command exterior devices like robotic limbs or computer borders. This article presents a Boosted Harris Hawks Shuffled Shepherd Optimization Augmented Deep Learning (BHHSHO-DL) technique based on Motor Imagery Classification for BCI. The BHHSHO-DL technique mainly exploits the hyperparameter-tuned DL approach for MI identification for BCI. Initially, the BHHSHO-DL technique performs data preprocessing utilizing the wavelet packet decomposition (WPD) model. Besides, the enhanced densely connected networks (DenseNet) model extracts the preprocessed data's complex and hierarchical feature patterns. Meanwhile, the BHHSHO technique-based hyperparameter tuning process is accomplished to elect optimal parameter values of the enhanced DenseNet model. Finally, the classification procedure is implemented by utilizing the convolutional autoencoder (CAE) model. The simulation value of the BHHSHO-DL methodology is performed on a benchmark dataset. The performance validation of the BHHSHO-DL methodology portrayed a superior accuracy value of 98.15% and 92.23% over other techniques under BCIC-III and BCIC-IV datasets.

PMID:39570847 | DOI:10.1371/journal.pone.0313261

J Chem Inf Model. 2024 Nov 21. doi: 10.1021/acs.jcim.4c01554. Online ahead of print.

ABSTRACT

Identifying drug-target interactions (DTIs) is essential for drug discovery and development. Existing deep learning approaches to DTI prediction often employ powerful feature encoders to represent drugs and targets holistically, which usually cause significant redundancy and noise by neglecting the restricted binding regions. Furthermore, many previous DTI networks ignore or simplify the complex intermolecular interaction process involving diverse binding types, which significantly limits both predictive ability and interpretability. We propose ReduMixDTI, an end-to-end model that addresses feature redundancy and explicitly captures complex local interactions for DTI prediction. In this study, drug and target features are encoded by using graph neural networks and convolutional neural networks, respectively. These features are refined from channel and spatial perspectives to enhance the representations. The proposed attention mechanism explicitly models pairwise interactions between drug and target substructures, improving the model's understanding of binding processes. In extensive comparisons with seven state-of-the-art methods, ReduMixDTI demonstrates superior performance across three benchmark data sets and external test sets reflecting real-world scenarios. Additionally, we perform comprehensive ablation studies and visualize protein attention weights to enhance the interpretability. The results confirm that ReduMixDTI serves as a robust and interpretable model for reducing feature redundancy, contributing to advances in DTI prediction.

PMID:39570771 | DOI:10.1021/acs.jcim.4c01554

Bioinformatics. 2024 Nov 21:btae675. doi: 10.1093/bioinformatics/btae675. Online ahead of print.

ABSTRACT

MOTIVATION: Post-translational modification (PTM) crosstalk events play critical roles in biological processes. Several machine learning methods have been developed to identify PTM crosstalk within proteins, but the accuracy is still far from satisfactory. Recent breakthroughs in deep learning and protein structure prediction could provide a potential solution to this issue.

RESULTS: We proposed DeepPCT, a deep learning algorithm to identify PTM crosstalk using AlphaFold2-based structures. In this algorithm, one deep learning classifier was constructed for sequence-based prediction by combining the residue and residue pair embeddings with cross-attention techniques, while the other classifier was established for structure-based prediction by integrating the structural embedding and a graph neural network. Meanwhile, a machine learning classifier was developed using novel structural descriptors and a random forest model to complement the structural deep learning classifier. By integrating the three classifiers, DeepPCT outperformed existing algorithms in different evaluation scenarios and showed better generalizability on new data owing to its less distance dependency.

AVAILABILITY: Datasets, codes, and models of DeepPCT are freely accessible at https://github.com/hzau-liulab/DeepPCT/.

SUPPLEMENTARY INFORMATION: Supplementary data are available at Bioinformatics online.

PMID:39570595 | DOI:10.1093/bioinformatics/btae675

Traffic Inj Prev. 2024 Nov 21:1-8. doi: 10.1080/15389588.2024.2402936. Online ahead of print.

ABSTRACT

OBJECTIVE: The objective of this study is to conduct precise risk prediction of traffic accidents using vehicle trajectory data.

METHODS: For urban road and highway scenarios, a scheme was developed to gather vehicle kinematic data and driving operation records from an in-vehicle device. The raw trajectory samples of over 3000 vehicles were processed through cleaning, filtering, interpolation, and normalization for preprocessing. Three deep learning frameworks based on RNN, CNN, and LSTM were compared. An end-to-end LSTM accident risk prediction model was constructed, and the model was trained using the cross-entropy loss function with Adam optimizer.

RESULTS: The LSTM model is capable of directly extracting accident-related hazardous state features from low-quality raw trajectory data, thereby enabling the prediction of accident probability with fine-grained time resolution. In tests conducted under complex traffic scenarios, the model successfully identifies high-risk driving behaviors in high-speed road sections and intersections with a prediction accuracy of 0.89, demonstrating strong generalization performance.

CONCLUSIONS: The LSTM accident risk prediction model, based on vehicle trajectory, developed in this study, is capable of intelligently extracting dangerous driving features. It can accurately warn about the risk of traffic accidents and provides a novel approach to enhancing road safety.

PMID:39570198 | DOI:10.1080/15389588.2024.2402936

SICOT J. 2024;10:49. doi: 10.1051/sicotj/2024044. Epub 2024 Nov 21.

ABSTRACT

The integration of artificial intelligence (AI) into orthopaedic care has gained considerable interest in recent years, evidenced by the growing body of literature boasting wide-ranging applications across the perioperative setting. This includes automated diagnostic imaging, clinical decision-making tools, optimisation of implant design, robotic surgery, and remote patient monitoring. Collectively, these advances propose to enhance patient care and improve system efficiency. Musculoskeletal pathologies represent the most significant contributor to global disability, with roughly 1.71 billion people afflicted, leading to an increasing volume of patients awaiting planned orthopaedic surgeries. This has exerted a considerable strain on healthcare systems globally, compounded by both the COVID-19 pandemic and the effects of an ageing population. Subsequently, patients face prolonged waiting times for surgery, with further deterioration and potentially poorer outcomes as a result. Furthermore, incorporating AI technologies into clinical practice could provide a means of addressing current and future service demands. This review aims to present a clear overview of AI applications across preoperative, intraoperative, and postoperative stages to elucidate its potential to transform planned orthopaedic care.

PMID:39570038 | DOI:10.1051/sicotj/2024044

Phys Med Biol. 2024 Nov 20. doi: 10.1088/1361-6560/ad953e. Online ahead of print.

ABSTRACT

The latest developments combining deep learning technology and medical image data have attracted wide attention and provide efficient noninvasive methods for the early diagnosis of breast cancer. The success of this task often depends on a large amount of data annotated by medical experts, which is time-consuming and may not always be feasible in the biomedical field. The lack of interpretability has greatly hindered the application of deep learning in the medical field. Currently, deep stable learning, including causal inference, make deep learning models more predictive and interpretable. In this study, to distinguish malignant tumors in Breast Imaging-Reporting and Data System (BI-RADS) category 3-4A breast lesions, we propose BD-StableNet, a deep stable learning model for the automatic detection of lesion areas. In this retrospective study, we collected 3103 breast ultrasound images (1418 benign and 1685 malignant lesions) from 493 patients (361 benign and 132 malignant lesion patients) for model training and testing. Compared with other mainstream deep learning models, BD-StableNet has better prediction performance (accuracy = 0.952, area under the curve (AUC) = 0.982, precision = 0.970, recall = 0.941, F1-score = 0.955 and specificity = 0.965). The lesion area prediction and class activation map (CAM) results both verify that our proposed model is highly interpretable. The results indicate that BD-StableNet significantly enhances diagnostic accuracy and interpretability, offering a promising noninvasive approach for the diagnosis of BI-RADS category 3-4A breast lesions. Clinically, the use of BD-StableNet could reduce unnecessary biopsies, improve diagnostic efficiency, and ultimately enhance patient outcomes by providing more precise and reliable assessments of breast lesions.

PMID:39569908 | DOI:10.1088/1361-6560/ad953e

J Neural Eng. 2024 Nov 19. doi: 10.1088/1741-2552/ad94a4. Online ahead of print.

ABSTRACT

OBJECTIVE: Visual prostheses are effective tools for restoring vision, yet real-world complexities pose ongoing challenges. The progress in AI has led to the emergence of the concept of intelligent visual prosthetics with auditory support, leveraging deep learning to create practical artificial vision perception beyond merely restoring natural sight for the blind.

APPROACH: This study introduces an object-based attention mechanism that simulates human gaze points when observing the external world to descriptions of physical regions. By transforming this mechanism into a ranking problem of salient entity regions, we introduce prior visual attention cues to build a new salient object ranking dataset, and propose a salient object ranking (SaOR) network aimed at providing depth perception for prosthetic vision. Furthermore, we propose a SaOR-guided image description method to align with human observation patterns, toward providing additional visual information by auditory feedback. Finally, the integration of the two aforementioned algorithms constitutes an audiovisual cognitive optimization strategy for prosthetic vision.

MAIN RESULTS: Through conducting psychophysical experiments based on scene description tasks under simulated prosthetic vision, we verify that the SaOR method improves the subjects' performance in terms of object identification and understanding the correlation among objects. Additionally, the cognitive optimization strategy incorporating image description further enhances their prosthetic visual cognition.

SIGNIFICANCE: This offers valuable technical insights for designing next-generation intelligent visual prostheses and establishes a theoretical groundwork for developing their visual information processing strategies. Code will be made publicly available.

PMID:39569905 | DOI:10.1088/1741-2552/ad94a4

Phys Med Biol. 2024 Nov 19. doi: 10.1088/1361-6560/ad94c7. Online ahead of print.

ABSTRACT

Magnetic resonance imaging (MRI) provides detailed structural information of the internal body organs and soft tissue regions of a patient in clinical diagnosis for disease detection, localization, and progress monitoring. MRI scanner hardware manufacturers incorporate various post-acquisition image-processing techniques into the scanner's computer software tools for different post-processing tasks. These tools provide a final image of adequate quality and essential features for accurate clinical reporting and predictive interpretation for better treatment planning. Different post-acquisition image-processing tasks for MRI quality enhancement include noise removal, motion artifact reduction, magnetic bias field correction, and eddy electric current effect removal. Recently, deep learning (DL) methods have shown great success in many research fields, including image and video applications. DL-based data-driven feature-learning approaches have great potential for MR image denoising and image-quality-degrading artifact correction. Recent studies have demonstrated significant improvements in image-analysis tasks using DL-based convolutional neural network (CNN) techniques. The promising capabilities and performance of DL techniques in various problem-solving domains have motivated researchers to adapt DL methods to medical image analysis and quality enhancement tasks. This paper presents a comprehensive review of DL-based state-of-the-art MRI quality enhancement and artifact removal methods for regenerating high-quality images while preserving essential anatomical and physiological feature maps without destroying important image information. Existing research gaps and future directions have also been provided by highlighting potential research areas for future developments, along with their importance and advantages in medical imaging.

PMID:39569887 | DOI:10.1088/1361-6560/ad94c7

Biomed Phys Eng Express. 2024 Nov 19. doi: 10.1088/2057-1976/ad947b. Online ahead of print.

ABSTRACT


This paper introduces a deep learning method for myocardial strain analysis while also evaluating the efficacy of the method across a public and a private dataset for cardiac pathology discrimination.
Methods:
We measure the global and regional myocardial strain in cSAX CMR images by first identifying a ROI centered in the LV, obtaining the cardiac structures (LV, RV and Myo) and estimating the motion of the myocardii. Finally we compute the strain for the heart coordinate system and report the global and regional strain.
Results:
We validated our method in two public datasets (ACDC, 80 subjects and CMAC, 16 subjects) and a private dataset (SSC, 75 subjects), containing healthy and pathological cases (acute myocardial infarct, DCM and HCM). We measured the mean Dice coefficient and Haussdorff distance for segmentation accuracy, the absolute end point error for motion accuracy, and we conducted a study of the discrimination power of the strain and strain rate between populations of healthy and pathological subjects. The results demonstrated that our method effectively quantifies myocardial strain and strain rate, showing distinct patterns across different cardiac conditions achieving notable statistical significance. Results also show that the method's accuracy is on par with iterative non-parametric registration methods and is also capable of estimating regional strain values.
Conclusion:
Our method proves to be a powerful tool for cardiac strain analysis, achieving results comparable to other state of the art methods, and computational efficiency over traditional methods.&#xD.

PMID:39569845 | DOI:10.1088/2057-1976/ad947b

Med Phys. 2024 Nov 21. doi: 10.1002/mp.17512. Online ahead of print.

ABSTRACT

BACKGROUND: PET/CT and planning CT are commonly used medical images in radiotherapy for esophageal and nasopharyngeal cancer. However, repeated scans will expose patients to additional radiation doses and also introduce registration errors. This multimodal treatment approach is expected to be further improved.

PURPOSE: A new Transformer model is proposed to obtain pseudo-PET/CT fusion images for esophageal and nasopharyngeal cancer radiotherapy.

METHODS: The data of 129 cases of esophageal cancer and 141 cases of nasopharyngeal cancer were retrospectively selected for training, validation, and testing. PET and CT images are used as input. Based on the Transformer model with a "focus-disperse" attention mechanism and multi-consistency loss constraints, the feature information in two images is effectively captured. This ultimately results in the synthesis of pseudo-PET/CT fusion images with enhanced tumor region imaging. During the testing phase, the accuracy of pseudo-PET/CT fusion images was verified in anatomy and dosimetry, and two prospective cases were selected for further dose verification.

RESULTS: In terms of anatomical verification, the PET/CT fusion image obtained using the wavelet fusion algorithm was used as the ground truth image after correction by clinicians. The evaluation metrics, including peak signal-to-noise ratio, structural similarity index, mean absolute error, and normalized root mean square error, between the pseudo-fused images obtained based on the proposed model and ground truth, are represented by means (standard deviation). They are 37.82 (1.57), 95.23 (2.60), 29.70 (2.49), and 9.48 (0.32), respectively. These numerical values outperform those of the state-of-the-art deep learning comparative models. In terms of dosimetry validation, based on a 3%/2 mm gamma analysis, the average passing rates of global and tumor regions between the pseudo-fused images (with a PET/CT weight ratio of 2:8) and the planning CT images are 97.2% and 95.5%, respectively. These numerical outcomes are superior to those of pseudo-PET/CT fusion images with other weight ratios.

CONCLUSIONS: This pseudo-PET/CT fusion images obtained based on the proposed model hold promise as a new modality in the radiotherapy for esophageal and nasopharyngeal cancer.

PMID:39569842 | DOI:10.1002/mp.17512

Med Phys. 2024 Nov 21. doi: 10.1002/mp.17521. Online ahead of print.

ABSTRACT

BACKGROUND: Mammographic imaging is essential for breast cancer detection and diagnosis. In addition to masses, calcifications are of concern and the early detection of breast cancer also heavily relies on the correct interpretation of suspicious microcalcification clusters. Even with advances in imaging and the introduction of novel techniques such as digital breast tomosynthesis and contrast-enhanced mammography, a correct interpretation can still be challenging given the subtle nature and large variety of calcifications.

PURPOSE: Computer simulated lesion models can serve to develop, optimize, or improve imaging techniques. In addition to their use in comparative (virtual clinical trial) detection experiments, these models have potential application in training deep learning models and in the understanding and interpretation of breast lesions. Existing simulation methods, however, often lack the capacity to model the diversity occurring in breast lesions or to generate models relevant for a specific case. This study focuses on clusters of microcalcifications and introduces an automated, flexible toolbox designed to generate microcalcification cluster models customized to specific tasks.

METHODS: The toolbox allows users to control a large number of simulation parameters related to model characteristics such as lesion size, calcification shape, or number of microcalcifications per cluster. This leads to the capability of creating models that range from regular to complex clusters. Based on the input parameters, which are either tuned manually or pre-set for a specific clinical type, different sets of models can be simulated depending on the use case. Two lesion generation methods are described. The first method generates three-dimensional microcalcification clusters models based on geometrical shapes and transformations. The second method creates two-dimensional (2D) microcalcification cluster models for a specific 2D mammographic image. This novel method employs radiomics analysis to account for local textures, ensuring the simulated microcalcification cluster is appropriately integrated within the existing breast tissue. The toolbox is implemented in the Python language and can be conveniently run through a Jupyter Notebook interface, openly accessible at https://gitlab.kuleuven.be/medphysqa/deploy/breast-calcifications. Validation studies performed by radiologists assessed the level of malignancy and realism of clusters tuned with specific parameters and inserted in mammographic images.

RESULTS: The flexibility of the toolbox with multiple simulation methods is illustrated, as well as the compatibility with different simulation frameworks and image types. The automation allows for the straightforward and fast generation of diverse microcalcification cluster models. The generated models are most likely applicable for various tasks as they can be configured in a variety of ways and inserted in different types of mammographic images of multiple acquisition systems. Validation studies confirmed the capacity to simulate realistic clusters and capture clinical properties when tuned with appropriate parameter settings.

CONCLUSION: This simulation toolbox offers a flexible means of simulating microcalcification cluster models with potential use in both technical and clinical research in mammography imaging. The 3D generation methods allow for specifying many characteristics regarding the calcification shape and cluster architecture, and the 2D generation method presents a novel manner to create microcalcification clusters tailored to existing breast textures.

PMID:39569820 | DOI:10.1002/mp.17521