RESUMO
Brain tumors occur due to the expansion of abnormal cell tissues and can be malignant (cancerous) or benign (not cancerous). Numerous factors such as the position, size, and progression rate are considered while detecting and diagnosing brain tumors. Detecting brain tumors in their initial phases is vital for diagnosis where MRI (magnetic resonance imaging) scans play an important role. Over the years, deep learning models have been extensively used for medical image processing. The current study primarily investigates the novel Fine-Tuned Vision Transformer models (FTVTs)-FTVT-b16, FTVT-b32, FTVT-l16, FTVT-l32-for brain tumor classification, while also comparing them with other established deep learning models such as ResNet50, MobileNet-V2, and EfficientNet - B0. A dataset with 7,023 images (MRI scans) categorized into four different classes, namely, glioma, meningioma, pituitary, and no tumor are used for classification. Further, the study presents a comparative analysis of these models including their accuracies and other evaluation metrics including recall, precision, and F1-score across each class. The deep learning models ResNet-50, EfficientNet-B0, and MobileNet-V2 obtained an accuracy of 96.5%, 95.1%, and 94.9%, respectively. Among all the FTVT models, FTVT-l16 model achieved a remarkable accuracy of 98.70% whereas other FTVT models FTVT-b16, FTVT-b32, and FTVT-132 achieved an accuracy of 98.09%, 96.87%, 98.62%, respectively, hence proving the efficacy and robustness of FTVT's in medical image processing.
RESUMO
Medical image segmentation is crucial for healthcare, yet convolution-based methods like U-Net face limitations in modeling long-range dependencies. To address this, Transformers designed for sequence-to-sequence predictions have been integrated into medical image segmentation. However, a comprehensive understanding of Transformers' self-attention in U-Net components is lacking. TransUNet, first introduced in 2021, is widely recognized as one of the first models to integrate Transformer into medical image analysis. In this study, we present the versatile framework of TransUNet that encapsulates Transformers' self-attention into two key modules: (1) a Transformer encoder tokenizing image patches from a convolution neural network (CNN) feature map, facilitating global context extraction, and (2) a Transformer decoder refining candidate regions through cross-attention between proposals and U-Net features. These modules can be flexibly inserted into the U-Net backbone, resulting in three configurations: Encoder-only, Decoder-only, and Encoder+Decoder. TransUNet provides a library encompassing both 2D and 3D implementations, enabling users to easily tailor the chosen architecture. Our findings highlight the encoder's efficacy in modeling interactions among multiple abdominal organs and the decoder's strength in handling small targets like tumors. It excels in diverse medical applications, such as multi-organ segmentation, pancreatic tumor segmentation, and hepatic vessel segmentation. Notably, our TransUNet achieves a significant average Dice improvement of 1.06% and 4.30% for multi-organ segmentation and pancreatic tumor segmentation, respectively, when compared to the highly competitive nn-UNet, and surpasses the top-1 solution in the BrasTS2021 challenge. 2D/3D Code and models are available at https://github.com/Beckschen/TransUNet and https://github.com/Beckschen/TransUNet-3D, respectively.
RESUMO
Efficient and accurate classification of DNA barcode data is crucial for large-scale fungal biodiversity studies. However, existing methods are either computationally expensive or lack accuracy. Previous research has demonstrated the potential of deep learning in this domain, successfully training neural networks for biological sequence classification. We introduce the MycoAI Python package, featuring various deep learning models such as BERT and CNN tailored for fungal Internal Transcribed Spacer (ITS) sequences. We explore different neural architecture designs and encoding methods to identify optimal models. By employing a multi-head output architecture and multi-level hierarchical label smoothing, MycoAI effectively generalizes across the taxonomic hierarchy. Using over 5 million labelled sequences from the UNITE database, we develop two models: MycoAI-BERT and MycoAI-CNN. While we emphasize the necessity of verifying classification results by AI models due to insufficient reference data, MycoAI still exhibits substantial potential. When benchmarked against existing classifiers such as DNABarcoder and RDP on two independent test sets with labels present in the training dataset, MycoAI models demonstrate high accuracy at the genus and higher taxonomic levels, with MycoAI-CNN being the fastest and most accurate. In terms of efficiency, MycoAI models can classify over 300,000 sequences within 5 min. We publicly release the MycoAI models, enabling mycologists to classify their ITS barcode data efficiently. Additionally, MycoAI serves as a platform for developing further deep learning-based classification methods. The source code for MycoAI is available under the MIT Licence at https://github.com/MycoAI/MycoAI.
RESUMO
Survival analysis is employed to scrutinize time-to-event data, with emphasis on comprehending the duration until the occurrence of a specific event. In this article, we introduce two novel survival prediction models: CosAttnSurv and CosAttnSurv + DyACT. CosAttnSurv model leverages transformer-based architecture and a softmax-free kernel attention mechanism for survival prediction. Our second model, CosAttnSurv + DyACT, enhances CosAttnSurv with Dynamic Adaptive Computation Time (DyACT) control, optimizing computation efficiency. The proposed models are validated using two public clinical datasets related to heart disease patients. When compared to other state-of-the-art models, our models demonstrated an enhanced discriminative and calibration performance. Furthermore, in comparison to other transformer architecture-based models, our proposed models demonstrate comparable performance while exhibiting significant reduction in both time and memory requirements. Overall, our models offer significant advancements in the field of survival analysis and emphasize the importance of computationally effective time-based predictions, with promising implications for medical decision-making and patient care.
RESUMO
In recent years, the field of artificial intelligence has witnessed a remarkable surge in the generation of synthetic images, driven by advancements in deep learning techniques. These synthetic images, often created through complex algorithms, closely mimic real photographs, blurring the lines between reality and artificiality. This proliferation of synthetic visuals presents a pressing challenge: how to accurately and reliably distinguish between genuine and generated images. This article, in particular, explores the task of detecting images generated by text-to-image diffusion models, highlighting the challenges and peculiarities of this field. To evaluate this, we consider images generated from captions in the MSCOCO and Wikimedia datasets using two state-of-the-art models: Stable Diffusion and GLIDE. Our experiments show that it is possible to detect the generated images using simple multi-layer perceptrons (MLPs), starting from features extracted by CLIP or RoBERTa, or using traditional convolutional neural networks (CNNs). These latter models achieve remarkable performances in particular when pretrained on large datasets. We also observe that models trained on images generated by Stable Diffusion can occasionally detect images generated by GLIDE, but only on the MSCOCO dataset. However, the reverse is not true. Lastly, we find that incorporating the associated textual information with the images in some cases can lead to a better generalization capability, especially if textual features are closely related to visual ones. We also discovered that the type of subject depicted in the image can significantly impact performance. This work provides insights into the feasibility of detecting generated images and has implications for security and privacy concerns in real-world applications. The code to reproduce our results is available at: https://github.com/davide-coccomini/Detecting-Images-Generated-by-Diffusers.
RESUMO
Radiologists encounter significant challenges when segmenting and determining brain tumors in patients because this information assists in treatment planning. The utilization of artificial intelligence (AI), especially deep learning (DL), has emerged as a useful tool in healthcare, aiding radiologists in their diagnostic processes. This empowers radiologists to understand the biology of tumors better and provide personalized care to patients with brain tumors. The segmentation of brain tumors using multi-modal magnetic resonance imaging (MRI) images has received considerable attention. In this survey, we first discuss multi-modal and available magnetic resonance imaging modalities and their properties. Subsequently, we discuss the most recent DL-based models for brain tumor segmentation using multi-modal MRI. We divide this section into three parts based on the architecture: the first is for models that use the backbone of convolutional neural networks (CNN), the second is for vision transformer-based models, and the third is for hybrid models that use both convolutional neural networks and transformer in the architecture. In addition, in-depth statistical analysis is performed of the recent publication, frequently used datasets, and evaluation metrics for segmentation tasks. Finally, open research challenges are identified and suggested promising future directions for brain tumor segmentation to improve diagnostic accuracy and treatment outcomes for patients with brain tumors. This aligns with public health goals to use health technologies for better healthcare delivery and population health management.
RESUMO
Accurate segmentation of chronic stroke lesions from mono-spectral magnetic resonance imaging scans (e.g., T1-weighted images) is a difficult task due to the arbitrary shape, complex texture, variable size and intensities, and varied locations of the lesions. Due to this inherent spatial heterogeneity, existing machine learning methods have shown moderate performance for chronic lesion delineation. In this study, we introduced: (1) a method that integrates transformers' deformable feature attention mechanism with convolutional deep learning architecture to improve the accuracy and generalizability of stroke lesion segmentation, and (2) an ecological data augmentation technique based on inserting real lesions into intact brain regions. Our combination of these two approaches resulted in a significant increase in segmentation performance, with a Dice index of 0.82 (±0.39), outperforming the existing methods trained and tested on the same Anatomical Tracings of Lesions After Stroke (ATLAS) 2022 dataset. Our method performed relatively well even for cases with small stroke lesions. We validated the robustness of our method through an ablation study and by testing it on new unseen brain scans from the Ischemic Stroke Lesion Segmentation (ISLES) 2015 dataset. Overall, our proposed approach of transformers with ecological data augmentation offers a robust way to delineate chronic stroke lesions with clinically relevant accuracy. Our method can be extended to other challenging tasks that require automated detection and segmentation of diverse brain abnormalities from clinical scans.
Assuntos
Aprendizado Profundo , Imageamento por Ressonância Magnética , Acidente Vascular Cerebral , Humanos , Imageamento por Ressonância Magnética/métodos , Imageamento por Ressonância Magnética/normas , Acidente Vascular Cerebral/diagnóstico por imagem , Acidente Vascular Cerebral/patologia , Neuroimagem/métodos , Neuroimagem/normas , AVC Isquêmico/diagnóstico por imagem , Processamento de Imagem Assistida por Computador/métodos , Idoso , Encéfalo/diagnóstico por imagem , Encéfalo/patologiaRESUMO
Ultrahigh-frequency (UHF) sensing is one of the most promising techniques for assessing the quality of power transformer insulation systems due to its capability to identify failures like partial discharges (PDs) by detecting the emitted UHF signals. However, there are still uncertainties regarding the frequency range that should be evaluated in measurements. For example, most publications have stated that UHF emissions range up to 3 GHz. However, a Cigré brochure revealed that the optimal spectrum is between 100 MHz and 1 GHz, and more recently, a study indicated that the optimal frequency range is between 400 MHz and 900 MHz. Since different faults require different maintenance actions, both science and industry have been developing systems that allow for failure-type identification. Hence, it is important to note that bandwidth reduction may impair classification systems, especially those that are frequency-based. This article combines three operational conditions of a power transformer (healthy state, electric arc failure, and partial discharges on bushing) with three different self-organized maps to carry out failure classification: the chromatic technique (CT), principal component analysis (PCA), and the shape analysis clustering technique (SACT). For each case, the frequency content of UHF signals was selected at three frequency bands: the full spectrum, Cigré brochure range, and between 400 MHz and 900 MHz. Therefore, the contributions of this work are to assess how spectrum band limitation may alter failure classification and to evaluate the effectiveness of signal processing methodologies based on the frequency content of UHF signals. Additionally, an advantage of this work is that it does not rely on training as is the case for some machine learning-based methods. The results indicate that the reduced frequency range was not a limiting factor for classifying the state of the operation condition of the power transformer. Therefore, there is the possibility of using lower frequency ranges, such as from 400 MHz to 900 MHz, contributing to the development of less costly data acquisition systems. Additionally, PCA was found to be the most promising technique despite the reduction in frequency band information.
RESUMO
Molecular Property Prediction (MPP) is vital for drug discovery, crop protection, and environmental science. Over the last decades, diverse computational techniques have been developed, from using simple physical and chemical properties and molecular fingerprints in statistical models and classical machine learning to advanced deep learning approaches. In this review, we aim to distill insights from current research on employing transformer models for MPP. We analyze the currently available models and explore key questions that arise when training and fine-tuning a transformer model for MPP. These questions encompass the choice and scale of the pretraining data, optimal architecture selections, and promising pretraining objectives. Our analysis highlights areas not yet covered in current research, inviting further exploration to enhance the field's understanding. Additionally, we address the challenges in comparing different models, emphasizing the need for standardized data splitting and robust statistical analysis.
RESUMO
Anticancer peptides (ACPs) are a class of molecules that have gained significant attention in the field of cancer research and therapy. ACPs are short chains of amino acids, the building blocks of proteins, and they possess the ability to selectively target and kill cancer cells. One of the key advantages of ACPs is their ability to selectively target cancer cells while sparing healthy cells to a greater extent. This selectivity is often attributed to differences in the surface properties of cancer cells compared to normal cells. That is why ACPs are being investigated as potential candidates for cancer therapy. ACPs may be used alone or in combination with other treatment modalities like chemotherapy and radiation therapy. While ACPs hold promise as a novel approach to cancer treatment, there are challenges to overcome, including optimizing their stability, improving selectivity, and enhancing their delivery to cancer cells, continuous increasing in number of peptide sequences, developing a reliable and precise prediction model. In this work, we propose an efficient transformer-based framework to identify ACPs for by performing accurate a reliable and precise prediction model. For this purpose, four different transformer models, namely ESM, ProtBERT, BioBERT, and SciBERT are employed to detect ACPs from amino acid sequences. To demonstrate the contribution of the proposed framework, extensive experiments are carried on widely-used datasets in the literature, two versions of AntiCp2, cACP-DeepGram, ACP-740. Experiment results show the usage of proposed model enhances classification accuracy when compared to the literature studies. The proposed framework, ESM, exhibits 96.45% of accuracy for AntiCp2 dataset, 97.66% of accuracy for cACP-DeepGram dataset, and 88.51% of accuracy for ACP-740 dataset, thence determining new state-of-the-art. The code of proposed framework is publicly available at github (https://github.com/mstf-yalcin/acp-esm).
RESUMO
Cancer research encompasses data across various scales, modalities, and resolutions, from screening and diagnostic imaging to digitized histopathology slides to various types of molecular data and clinical records. The integration of these diverse data types for personalized cancer care and predictive modeling holds the promise of enhancing the accuracy and reliability of cancer screening, diagnosis, and treatment. Traditional analytical methods, which often focus on isolated or unimodal information, fall short of capturing the complex and heterogeneous nature of cancer data. The advent of deep neural networks has spurred the development of sophisticated multimodal data fusion techniques capable of extracting and synthesizing information from disparate sources. Among these, Graph Neural Networks (GNNs) and Transformers have emerged as powerful tools for multimodal learning, demonstrating significant success. This review presents the foundational principles of multimodal learning including oncology data modalities, taxonomy of multimodal learning, and fusion strategies. We delve into the recent advancements in GNNs and Transformers for the fusion of multimodal data in oncology, spotlighting key studies and their pivotal findings. We discuss the unique challenges of multimodal learning, such as data heterogeneity and integration complexities, alongside the opportunities it presents for a more nuanced and comprehensive understanding of cancer. Finally, we present some of the latest comprehensive multimodal pan-cancer data sources. By surveying the landscape of multimodal data integration in oncology, our goal is to underline the transformative potential of multimodal GNNs and Transformers. Through technological advancements and the methodological innovations presented in this review, we aim to chart a course for future research in this promising field. This review may be the first that highlights the current state of multimodal modeling applications in cancer using GNNs and transformers, presents comprehensive multimodal oncology data sources, and sets the stage for multimodal evolution, encouraging further exploration and development in personalized cancer care.
RESUMO
With the objective of extracting new knowledge about rare diseases from social media messages, we evaluated three models on a Named Entity Recognition (NER) task, consisting of extracting phenotypes and treatments from social media messages. We trained the three models on a dataset with social media messages about Developmental and Epileptic Encephalopathies and more common diseases. This preliminary study revealed that CamemBERT and CamemBERT-bio exhibit similar performance on social media testimonials, slightly outperforming DrBERT. It also highlighted that their performance was lower on this type of data than on structured health datasets. Limitations, including a narrow focus on NER performance and dataset-specific evaluation, call for further research to fully assess model capabilities on larger and more diverse datasets.
Assuntos
Mídias Sociais , França , Humanos , Processamento de Linguagem Natural , Mineração de Dados/métodos , Doenças RarasRESUMO
In data assimilation, observations are fused with simulations to obtain an accurate estimate of the state and parameters for a given physical system. Combining data with a model, however, while accurately estimating uncertainty, is computationally expensive and infeasible to run in real-time for complex systems. Here, we present a novel particle filter methodology, the Deep Latent Space Particle filter or D-LSPF, that uses neural network-based surrogate models to overcome this computational challenge. The D-LSPF enables filtering in the low-dimensional latent space obtained using Wasserstein AEs with modified vision transformer layers for dimensionality reduction and transformers for parameterized latent space time stepping. As we demonstrate on three test cases, including leak localization in multi-phase pipe flow and seabed identification for fully nonlinear water waves, the D-LSPF runs orders of magnitude faster than a high-fidelity particle filter and 3-5 times faster than alternative methods while being up to an order of magnitude more accurate. The D-LSPF thus enables real-time data assimilation with uncertainty quantification for the test cases demonstrated in this paper.
RESUMO
The rise of electric vehicles (EVs) necessitates an efficient charging infrastructure capable of delivering a refueling experience akin to conventional vehicles. Innovations in Extreme Fast Charging (EFC) offer promising solutions in this regard. By harnessing renewable energy sources and employing sophisticated multiport converters, EFC systems can meet the evolving demands of EV refueling. A single-stage topology simplifies the converter design, focusing on efficient DC-AC conversion, vital for feeding solar power into the grid or charging stations. It provides power factor correction, harmonics filtering, and mitigates power quality issues, ensuring stable and efficient operations. Converters with Maximum Power Point Tracking (MPPT) capability facilitate the efficient integration of solar PV systems in charging stations, ensuring maximum solar energy utilization for EV charging. The ability to operate in different modes allows seamless integration with energy storage systems, storing excess solar energy for use during night-time or peak demand periods, enhancing overall efficiency and reliability. Advanced converters support bidirectional energy flow, enabling EV batteries to discharge back to the grid, aiding grid stability and energy management. However, robust control algorithms are needed to handle dynamic conditions like partial shading more effectively. Our review focuses on integrating renewable energy sources with multiport converters, providing insights into a novel EV charging station framework optimized for EFC topology. We highlight the advantages of multiport non-isolated converters over traditional line frequency transformers, particularly in medium voltage scenarios, offering enhanced efficiency and versatility for EFC applications.
RESUMO
Pneumonia is the leading cause of death among children around the world. According to WHO, a total of 740,180 lives under the age of five were lost due to pneumonia in 2019. Lung ultrasound (LUS) has been shown to be particularly useful for supporting the diagnosis of pneumonia in children and reducing mortality in resource-limited settings. The wide application of point-of-care ultrasound at the bedside is limited mainly due to a lack of training for data acquisition and interpretation. Artificial Intelligence can serve as a potential tool to automate and improve the LUS data interpretation process, which mainly involves analysis of hyper-echoic horizontal and vertical artifacts, and hypo-echoic small to large consolidations. This paper presents, Fused Lung Ultrasound Encoding-based Transformer (FLUEnT), a novel pediatric LUS video scoring framework for detecting lung consolidations using fused LUS encodings. Frame-level embeddings from a variational autoencoder, features from a spatially attentive ResNet-18, and encoded patient information as metadata combiningly form the fused encodings. These encodings are then passed on to the transformer for binary classification of the presence or absence of consolidations in the video. The video-level analysis using fused encodings resulted in a mean balanced accuracy of 89.3 %, giving an average improvement of 4.7 % points in comparison to when using these encodings individually. In conclusion, outperforming the state-of-the-art models by an average margin of 8 % points, our proposed FLUEnT framework serves as a benchmark for detecting lung consolidations in LUS videos from pediatric pneumonia patients.
RESUMO
On average, more than 5 million patients are admitted to intensive care units (ICUs) in the US, with mortality rates ranging from 10 to 29%. The acuity state of patients in the ICU can quickly change from stable to unstable, sometimes leading to life-threatening conditions. Early detection of deteriorating conditions can assist in more timely interventions and improved survival rates. While Artificial Intelligence (AI)-based models show potential for assessing acuity in a more granular and automated manner, they typically use mortality as a proxy of acuity in the ICU. Furthermore, these methods do not determine the acuity state of a patient (i.e., stable or unstable), the transition between acuity states, or the need for life-sustaining therapies. In this study, we propose APRICOT-M (Acuity Prediction in Intensive Care Unit-Mamba), a 1M-parameter state space-based neural network to predict acuity state, transitions, and the need for life-sustaining therapies in real-time among ICU patients. The model integrates ICU data in the preceding four hours (including vital signs, laboratory results, assessment scores, and medications) and patient characteristics (age, sex, race, and comorbidities) to predict the acuity outcomes in the next four hours. Our state space-based model can process sparse and irregularly sampled data without manual imputation, thus reducing the noise in input data and increasing inference speed. The model was trained on data from 107,473 patients (142,062 ICU admissions) from 55 hospitals between 2014-2017 and validated externally on data from 74,901 patients (101,356 ICU admissions) from 143 hospitals. Additionally, it was validated temporally on data from 12,927 patients (15,940 ICU admissions) from one hospital in 2018-2019 and prospectively on data from 215 patients (369 ICU admissions) from one hospital in 2021-2023. Three datasets were used for training and evaluation: the University of Florida Health (UFH) dataset, the electronic ICU Collaborative Research Database (eICU), and the Medical Information Mart for Intensive Care (MIMIC)-IV dataset. APRICOT-M significantly outperforms the baseline acuity assessment, Sequential Organ Failure Assessment (SOFA), for mortality prediction in both external (AUROC 0.95 CI: 0.94-0.95 compared to 0.78 CI: 0.78-0.79) and prospective (AUROC 0.99 CI: 0.97-1.00 compared to 0.80 CI: 0.65-0.92) cohorts, as well as for instability prediction (external AUROC 0.75 CI: 0.74-0.75 compared to 0.51 CI: 0.51-0.51, and prospective AUROC 0.69 CI: 0.64-0.74 compared to 0.53 CI: 0.50-0.57). This tool has the potential to help clinicians make timely interventions by predicting the transition between acuity states and decision-making on life-sustaining within the next four hours in the ICU.
RESUMO
BACKGROUND: The rapid growth of web-based medical services has highlighted the significance of smart triage systems in helping patients find the most appropriate physicians. However, traditional triage methods often rely on department recommendations and are insufficient to accurately match patients' textual questions with physicians' specialties. Therefore, there is an urgent need to develop algorithms for recommending physicians. OBJECTIVE: This study aims to develop and validate a patient-physician hybrid recommendation (PPHR) model with response metrics for better triage performance. METHODS: A total of 646,383 web-based medical consultation records from the Internet Hospital of the First Affiliated Hospital of Xiamen University were collected. Semantic features representing patients and physicians were developed to identify the set of most similar questions and semantically expand the pool of recommended physician candidates, respectively. The physicians' response rate feature was designed to improve candidate rankings. These 3 characteristics combine to create the PPHR model. Overall, 5 physicians participated in the evaluation of the efficiency of the PPHR model through multiple metrics and questionnaires as well as the performance of Sentence Bidirectional Encoder Representations from Transformers and Doc2Vec in text embedding. RESULTS: The PPHR model reaches the best recommendation performance when the number of recommended physicians is 14. At this point, the model has an F1-score of 76.25%, a proportion of high-quality services of 41.05%, and a rating of 3.90. After removing physicians' characteristics and response rates from the PPHR model, the F1-score decreased by 12.05%, the proportion of high-quality services fell by 10.87%, the average hit ratio dropped by 1.06%, and the rating declined by 11.43%. According to whether those 5 physicians were recommended by the PPHR model, Sentence Bidirectional Encoder Representations from Transformers achieved an average hit ratio of 88.6%, while Doc2Vec achieved an average hit ratio of 53.4%. CONCLUSIONS: The PPHR model uses semantic features and response metrics to enable patients to accurately find the physician who best suits their needs.
Assuntos
Médicos , Semântica , Humanos , Triagem/métodos , Triagem/normas , Inquéritos e Questionários , AlgoritmosRESUMO
The classification of carambola, also known as starfruit, according to quality parameters is usually conducted by trained human evaluators through visual inspections. This is a costly and subjective method that can generate high variability in results. As an alternative, computer vision systems (CVS) combined with deep learning (DCVS) techniques have been introduced in the industry as a powerful and an innovative tool for the rapid and non-invasive classification of fruits. However, validating the learning capability and trustworthiness of a DL model, aka black box, to obtain insights can be challenging. To reduce this gap, we propose an integrated eXplainable Artificial Intelligence (XAI) method for the classification of carambolas at different maturity stages. We compared two Residual Neural Networks (ResNet) and Visual Transformers (ViT) to identify the image regions that are enhanced by a Random Forest (RF) model, with the aim of providing more detailed information at the feature level for classifying the maturity stage. Changes in fruit colour and physicochemical data throughout the maturity stages were analysed, and the influence of these parameters on the maturity stages was evaluated using the Gradient-weighted Class Activation Mapping (Grad-CAM), the Attention Maps using RF importance. The proposed approach provides a visualization and description of the most important regions that led to the model decision, in wide visualization follows the models an importance features from RF. Our approach has promising potential for standardized and rapid carambolas classification, achieving 91 % accuracy with ResNet and 95 % with ViT, with potential application for other fruits.
Assuntos
Averrhoa , Frutas , Redes Neurais de Computação , Frutas/crescimento & desenvolvimento , Frutas/classificação , Averrhoa/química , Aprendizado Profundo , Inteligência Artificial , CorRESUMO
Background: Dual-energy CT (DECT) is a non-invasive way to determine the presence of monosodium urate (MSU) crystals in the workup of gout. Color-coding distinguishes MSU from calcium following material decomposition and post-processing. Most software labels MSU as green and calcium as blue. There are limitations in the current image processing methods of segmenting green-encoded pixels. Additionally, identifying green foci is tedious, and automated detection would improve workflow. This study aimed to determine the optimal deep learning (DL) algorithm for segmenting green-encoded pixels of MSU crystals on DECTs. Methods: DECT images of positive and negative gout cases were retrospectively collected. The dataset was split into train (N = 28) and held-out test (N = 30) sets. To perform cross-validation, the train set was split into seven folds. The images were presented to two musculoskeletal radiologists, who independently identified green-encoded voxels. Two 3D Unet-based DL models, Segresnet and SwinUNETR, were trained, and the Dice similarity coefficient (DSC), sensitivity, and specificity were reported as the segmentation metrics. Results: Segresnet showed superior performance, achieving a DSC of 0.9999 for the background pixels, 0.7868 for the green pixels, and an average DSC of 0.8934 for both types of pixels, respectively. According to the post-processed results, the Segresnet reached voxel-level sensitivity and specificity of 98.72 % and 99.98 %, respectively. Conclusion: In this study, we compared two DL-based segmentation approaches for detecting MSU deposits in a DECT dataset. The Segresnet resulted in superior performance metrics. The developed algorithm provides a potential fast, consistent, highly sensitive and specific computer-aided diagnosis tool. Ultimately, such an algorithm could be used by radiologists to streamline DECT workflow and improve accuracy in the detection of gout.
RESUMO
convolutional neural networks (CNNs) show great potential in medical image segmentation tasks, and can provide reliable basis for disease diagnosis and clinical research. However, CNNs exhibit general limitations on modeling explicit long-range relation, and existing cures, resorting to building deep encoders along with aggressive downsampling operations, leads to loss of localized details. Transformer has naturally excellent ability to model the global features and long-range correlations of the input information, which is strongly complementary to the inductive bias of CNNs. In this paper, a novel Bi-directional Multi-scale Cascaded Segmentation Network, BMCS-Net, is proposed to improve the performance of medical segmentation tasks by aggregating these features obtained from Transformers and CNNs branches. Specifically, a novel feature integration technique, termed as Two-stream Cascaded Feature Aggregation (TCFA) module, is designed to fuse features in two-stream branches, and solve the problem of gradual dilution of global information in the network. Besides, a Multi-Scale Expansion-Aware (MSEA) module based on the convolution of feature perception and expansion is introduced to capture context information, and further compensate for the loss of details. Extensive experiments demonstrated that BMCS-Net has an excellent performance on both skin and Polyp segmentation datasets.