RESUMEN
RATIONALE AND OBJECTIVES: To evaluate the performance of deep learning (DL) in predicting different breast cancer molecular subtypes using DCE-MRI from two institutes. MATERIALS AND METHODS: This retrospective study included 366 breast cancer patients from two institutes, divided into training (nâ¯=â¯292), validation (nâ¯=â¯49) and testing (nâ¯=â¯25) sets. We first transformed the public DCE-MRI appearance to ours to alleviate small-data-size and class-imbalance issues. Second, we developed a multi-branch convolutional-neural-network (MBCNN) to perform molecular subtype prediction. Third, we assessed the MBCNN with different regions of interest (ROIs) and fusion strategies, and compared it to previous DL models. Area under the curve (AUC) and accuracy (ACC) were used to assess different models. Delong-test was used for the comparison of different groups. RESULTS: MBCNN achieved the optimal performance under intermediate fusion and ROI size of 80 pixels with appearance transformation. It outperformed CNN and convolutional long-short-term-memory (CLSTM) in predicting luminal B, HER2-enriched and TN subtypes, but without demonstrating statistical significance except against CNN in TN subtypes, with testing AUCs of 0.8182 vs. [0.7208, 0.7922] (p=0.44, 0.80), 0.8500 vs. [0.7300, 0.8200] (p=0.36, 0.70) and 0.8900 vs. [0.7600, 0.8300] (p=0.03, 0.63), respectively. When predicting luminal A, MBCNN outperformed CNN with AUCs of 0.8571 vs. 0.7619 (p=0.08) without achieving statistical significance, and is comparable to CLSTM. For four-subtype prediction, MBCNN achieved an ACC of 0.64, better than CNN and CLSTM models with ACCs of 0.48 and 0.52, respectively. CONCLUSION: Developed DL model with the feature extraction and fusion of DCE-MRI from two institutes enabled preoperative prediction of breast cancer molecular subtypes with high diagnostic performance.
RESUMEN
Chemical exchange saturation transfer magnetic resonance imaging (CEST MRI) is a promising molecular imaging tool that allows sensitive detection of endogenous metabolic changes. However, because the CEST spectrum does not display a clear peak like MR spectroscopy, its signal interpretation is challenging, especially under 3-T field strength or with a large saturation B1 . Herein, as an alternative to conventional Z-spectral fitting approaches, a permuted random forest (PRF) method is developed to determine featured saturation frequencies for lesion identification, so-called CEST frequency importance analysis. Briefly, voxels in the CEST dataset were labeled as lesion and control according to multicontrast MR images. Then, by considering each voxel's saturation signal series as a sample, a permutation importance algorithm was employed to rank the contribution of saturation frequency offsets in the differentiation of lesion and normal tissue. Simulations demonstrated that PRF could correctly determine the frequency offsets (3.5 or -3.5 ppm) for classifying two groups of Z-spectra, under a range of B0 , B1 conditions and sample sizes. For ischemic rat brains, PRF only displayed high feature importance around amide frequency at 2 h postischemia, reflecting that the pH changes occurred at an early stage. By contrast, the data acquired at 24 h postischemia exhibited high feature importance at multiple frequencies (amide, water, and lipids), which suggested the complex tissue changes that occur during the later stages. Finally, PRF was assessed using 3-T CEST data from four brain tumor patients. By defining the tumor region on amide proton transfer-weighted images, PRF analysis identified different CEST frequency importance for two types of tumors (glioblastoma and metastatic tumor) (p < 0.05, with each image slice as a subject). In conclusion, the PRF method was able to rank and interpret the contribution of all acquired saturation offsets to lesion identification; this may facilitate CEST analysis in clinical applications, and open up new doors for comprehensive CEST analysis tools other than model-based approaches.
Asunto(s)
Neoplasias Encefálicas , Bosques Aleatorios , Ratas , Animales , Imagen por Resonancia Magnética/métodos , Neoplasias Encefálicas/diagnóstico por imagen , Neoplasias Encefálicas/metabolismo , Espectroscopía de Resonancia Magnética/métodos , Protones , AmidasRESUMEN
Objective. Efficient non-contact heart rate (HR) measurement from facial video has received much attention in health monitoring. Past methods relied on prior knowledge and an unproven hypothesis to extract remote photoplethysmography (rPPG) signals, e.g. manually designed regions of interest (ROIs) and the skin reflection model.Approach. This paper presents a short-time end to end HR estimation framework based on facial features and temporal relationships of video frames. In the proposed method, a deep 3D multi-scale network with cross-layer residual structure is designed to construct an autoencoder and extract robust rPPG features. Then, a spatial-temporal fusion mechanism is proposed to help the network focus on features related to rPPG signals. Both shallow and fused 3D spatial-temporal features are distilled to suppress redundant information in the complex environment. Finally, a data augmentation strategy is presented to solve the problem of uneven distribution of HR in existing datasets.Main results. The experimental results on four face-rPPG datasets show that our method overperforms the state-of-the-art methods and requires fewer video frames. Compared with the previous best results, the proposed method improves the root mean square error (RMSE) by 5.9%, 3.4% and 21.4% on the OBF dataset (intra-test), COHFACE dataset (intra-test) and UBFC dataset (cross-test), respectively.Significance. Our method achieves good results on diverse datasets (i.e. highly compressed video, low-resolution and illumination variation), demonstrating that our method can extract stable rPPG signals in short time.
Asunto(s)
Aprendizaje Profundo , Fotopletismografía , Fotopletismografía/métodos , Algoritmos , Frecuencia Cardíaca , Proyectos de InvestigaciónRESUMEN
The existing publicly available datasets with pixel-level labels contain limited categories, and it is difficult to generalize to the real world containing thousands of categories. In this paper, we propose an approach to generate object masks with detailed pixel-level structures/boundaries automatically to enable semantic image segmentation of thousands of targets in the real world without manually labelling. A Guided Filter Network (GFN) is first developed to learn the segmentation knowledge from an existed dataset, and such GFN then transfers the learned segmentation knowledge to generate initial coarse object masks for the target images. These coarse object masks are treated as pseudo labels to self-optimize the GFN iteratively in the target images. Our experiments on six image sets have demonstrated that our proposed approach can generate object masks with detailed pixel-level structures/boundaries, whose quality is comparable to the manually-labelled ones. Our proposed approach also achieves better performance on semantic image segmentation than most existing weakly-supervised, semi-supervised, and domain adaptation approaches under the same experimental conditions.
RESUMEN
Multi-keyword query is widely supported in text search engines. However, an analogue in image retrieval systems, multi-object query, is rarely studied. Meanwhile, traditional object-based image retrieval methods often involve multiple steps separately. In this work, we propose a weakly-supervised Deep Multiple Instance Hashing (DMIH) approach for multi-object image retrieval. Our DMIH approach, which leverages a popular CNN model to build the end-to-end relation between a raw image and the binary hash codes of its multiple objects, can support multi-object queries effectively and integrate object detection with hashing learning seamlessly. We treat object detection as a binary multiple instance learning (MIL) problem and such instances are automatically extracted from multi-scale convolutional feature maps. We also design a conditional random field (CRF) module to capture both the semantic and spatial relations among different class labels. For hashing training, we sample image pairs to learn their semantic relationships in terms of hash codes of the most probable proposals for owned labels as guided by object predictors. The two objectives benefit each other in a multi-task learning scheme. Finally, a two-level inverted index method is proposed to further speed up the retrieval of multi-object queries. Our DMIH approach outperforms state-of-the-arts on public benchmarks for object-based image retrieval and achieves promising results for multi-object queries.
RESUMEN
Directional modulation (DM), as an emerging promising physical layer security (PLS) technique at the transmitter side with the help of an antenna array, has developed rapidly over decades. In this study, a DM technique using a polarization sensitive array (PSA) to produce the modulation with different polarization states (PSs) at different directions is investigated. A PSA, as a vector sensor, can be employed for more effective DM for an additional degree of freedom (DOF) provided in the polarization domain. The polarization information can be exploited to transmit different data streams simultaneously at the same directions, same frequency, but with different PSs in the desired directions to increase the channel capacity, and with random PSs off the desired directions to enhance PLS. The proposed method has the capability of concurrently projecting independent signals into different specified spatial directions while simultaneously distorting signal constellation in all other directions. The symbol error rate (SER), secrecy rate, and the robustness of the proposed DM scheme are analyzed. Design examples for single- and multi-beam DM systems are also presented. Simulations corroborate that (1) the proposed method is more effective for PLS; (2) the proposed DM scheme is more power-efficient than the traditional artificial noise aided DM schemes; and (3) the channel capacity is significantly improved compared with conventional scalar antenna arrays.
RESUMEN
Directional modulation (DM) technique has the ability to enhance the physical layer security (PLS) of wireless communications. Conventional DM schemes are usually based on a single antenna array with the basic assumption that eavesdroppers (Eves) and legitimate users (LUs) are in different directions. However, it is possible that Eves are in the same direction as LUs in practical applications. As a result, signals received by Eves will be approximately the same or even in better quality than those received by LUs. To address the neighbor security issue, we introduce a multiple antenna arrays model at the transmitter side with the help of the artificial noise (AN)-aided DM technique to achieve secure and precise DM transmission in this paper. Meanwhile, to recover the mixed useful signals, two novel DM schemes based on single- and multi-carrier multiple antenna arrays model are proposed, respectively. In addition, the symbol error rate (SER), secrecy rate, and robustness performance of the proposed DM schemes were analyzed and simulated. Simulations validate the effectiveness of the proposed DM schemes and demonstrate that multiple antenna arrays model based DM methods outperform single antenna array model aided DM methods in security.
RESUMEN
Here, we investigate the security of the practical one-way CVQKD and CV-MDI-QKD systems under laser seeding attack. In particular, Eve can inject a suitable light into the laser diodes of the light source modules in the two kinds of practical CVQKD systems, which results in the increased intensity of the generated optical signal. The parameter estimation under laser seeding attack shows that the secret key rates of these two schemes may be overestimated, which indicates that this attack can open a security loophole for Eve to successfully obtain information about secret key in these practical CVQKD systems. To close this loophole, we propose a real-time monitoring scheme to precisely evaluate the secret key rates of these schemes. The analysis results indicate the implementation of the proposed monitoring scheme can effectively resist this potential attack.
RESUMEN
In the process of oil exploitation, the water level of an oil well can be predicted and the position of reservoir can be estimated by measuring the water content of crude oil, with reference for the automatic production of high efficiency in the oil field. In this paper, a visual measuring method for water content of crude oil is proposed. The oil and water in crude oil samples were separated by centrifugation, distillation or electric dehydration, and a water-oil layered mixture was formed according to the unequal densities. Then the volume ratio of water and oil was analyzed by digital image processing, and the water content and oil content was able to be calculated. A new method for measuring water content of crude oil based on IGAVD (image grayscale accumulated value difference) is proposed, which overcomes the uncertainty caused by environmental illumination and improves the measurement accuracy. In order to verify the effectiveness of the algorithm, a miniaturization and low-cost system prototype was developed. The experimental results show that the average power consumption is about 165 mW and the measuring error is less than 1.0%. At the same time, the real-time and remote transmission about measurement results can be realized.
RESUMEN
Fluorescence molecular tomography (FMT) has been a promising imaging tool because it allows an accurate localizaton and quantitative analysis of the fluorophore distribution in animals. It, however, is still a challenge since its reconstruction suffers from severe ill-posedness. This paper introduces a reconstruction frame based on three-way decisions (TWD) for the inverse problem of FMT. On the first stage, a reconstruction result on the whole region is obtained by a certain reconstruction algorithm. With TWD, the recovered result has been divided into three regions: fluorescent target region, boundary region, and background region. On the second stage, the boundary region and fluorescent target region have been combined into the permissible region of the target. Then a new reconstruction on the permissible region has been carried out and a new recovered result is obtained. With TWD again, the new result has been classified into three pairwise disjoint regions. And the new fluorescent target region is the final reconstructed result. Both numerical simulation experiments and a real mouse experiment are carried out to validate the feasibility and potential of the presented reconstruction frame. The results indicate that the proposed reconstuction strategy based on TWD can provide a good performance in FMT reconstruction.
Asunto(s)
Fluorescencia , Procesamiento de Imagen Asistido por Computador/métodos , Tomografía , Algoritmos , Modelos TeóricosRESUMEN
The recommender system is widely used in the field of e-commerce and plays an important role in guiding customers to make smart decisions. Although many algorithms are available in the recommender system, collaborative filtering is still one of the most used and successful recommendation technologies. In collaborative filtering, similarity calculation is the main issue. In order to improve the accuracy and quality of recommendations, we proposed an improved similarity model, which takes three impact factors of similarity into account to minimize the deviation of similarity calculation. Compared with the traditional similarity measure, the advantages of our proposed model are that it makes full use of rating data and solves the problem of co-rated items. To validate the efficiency of the proposed algorithm, experiments were performed on four datasets. Results show that the proposed method can effectively improve the preferences of the recommender system and it is suitable for the sparsity data.
Asunto(s)
Algoritmos , Comercio/estadística & datos numéricos , Internet , Conducta de Elección , Bases de Datos Factuales , Humanos , Difusión de la Información , Tecnología de la Información/estadística & datos numéricos , Modelos EstadísticosRESUMEN
This paper proposes a post-processing strategy for recovered images of fluorescence molecular tomography. A threshold value is adaptively obtained from the recovered images without external interference, which is objective because it is extracted from the reconstructed result. The recovered images from simulation experiments and physical phantom experiments are processed by this threshold method. And by visualization, the processed images are clearer than those with no post-processing. The full width at half-maximum and contrast-to-noise ratio are then utilized to further verify the effectiveness of the post-processing method, being capable of removing spurious information from the original images, thus bringing convenience to users.
RESUMEN
How to guarantee the practical security of continuous-variable quantum key distribution (CVQKD) system has been an important issue in the quantum cryptography applications. In contrast to the previous practical security strategies, which focus on the intercept-resend attack or the Gaussian attack, we investigate the practical security strategy based on a general attack, i.e., an arbitrated individual attack or collective attack on the system by Eve in this paper. The low bound of intensity disturbance of the local oscillator signal for eavesdropper successfully concealing herself is obtained, considering all noises can be used by Eve in the practical environment. Furthermore, we obtain an optimal monitoring condition for the practical CVQKD system so that legitimate communicators can monitor the general attack in real-time. As examples, practical security of two special systems, i.e., the Gaussian modulated coherent state CVQKD system and the middle-based CVQKD system, are investigated under the intercept-resend attacks.
RESUMEN
In this paper, a hierarchical deep multi-task learning (HD-MTL) algorithm is developed to support large-scale visual recognition (e.g., recognizing thousands or even tens of thousands of atomic object classes automatically). First, multiple sets of multi-level deep features are extracted from different layers of deep convolutional neural networks (deep CNNs), and they are used to achieve more effective accomplishment of the coarseto- fine tasks for hierarchical visual recognition. A visual tree is then learned by assigning the visually-similar atomic object classes with similar learning complexities into the same group, which can provide a good environment for determining the interrelated learning tasks automatically. By leveraging the inter-task relatedness (inter-class similarities) to learn more discriminative group-specific deep representations, our deep multi-task learning algorithm can train more discriminative node classifiers for distinguishing the visually-similar atomic object classes effectively. Our hierarchical deep multi-task learning (HD-MTL) algorithm can integrate two discriminative regularization terms to control the inter-level error propagation effectively, and it can provide an end-to-end approach for jointly learning more representative deep CNNs (for image representation) and more discriminative tree classifier (for large-scale visual recognition) and updating them simultaneously. Our incremental deep learning algorithms can effectively adapt both the deep CNNs and the tree classifier to the new training images and the new object classes. Our experimental results have demonstrated that our HD-MTL algorithm can achieve very competitive results on improving the accuracy rates for large-scale visual recognition.
RESUMEN
Limited-projection fluorescence molecular tomography (FMT) has short data acquisition time that allows fast resolving of the three-dimensional visualization of fluorophore within small animal in vivo. However, limited-projection FMT reconstruction suffers from severe ill-posedness because only limited projections are used for reconstruction. To alleviate the ill-posedness, a feasible region extraction strategy based on a double mesh is presented for limited-projection FMT. First, an initial result is rapidly recovered using a coarse discretization mesh. Then, the reconstructed fluorophore area in the initial result is selected as a feasible region to guide the reconstruction using a fine discretization mesh. Simulation experiments on a digital mouse and small animal experiment in vivo are performed to validate the proposed strategy. It demonstrates that the presented strategy provides a good distribution of fluorophore with limited projections of fluorescence measurements. Hence, it is suitable for reconstruction of limited-projection FMT.
Asunto(s)
Procesamiento de Imagen Asistido por Computador/métodos , Imagen Molecular/métodos , Tomografía Óptica/métodos , Animales , Simulación por Computador , Fluorescencia , Imagenología Tridimensional/métodos , Modelos Lineales , Ratones , Ratones Endogámicos BALB C , Modelos Biológicos , Fenómenos Ópticos , FotonesRESUMEN
Imaging at low light levels has drawn much attention. In this paper, a method is experimentally demonstrated to realize computational imaging under weak illumination conditions. In our experiment, only one single-photon detector was used to capture the photons. With the time-correlated single-photon-counting technique, photons at a quite low level can be recorded and the time distribution histograms were constructed. The intensity of the light can be estimated from the histograms. The detection model was discussed, and clear images were obtained through a ghost-imaging algorithm. In addition, we propose a modified algorithm for the conventional ghost-imaging method that works more efficiently than the traditional ghost-imaging algorithm. Moreover, this method provides a solution for three-dimensional imaging combining with the time of flight of the photons.
Asunto(s)
Algoritmos , Interpretación de Imagen Asistida por Computador/instrumentación , Imagenología Tridimensional/instrumentación , Iluminación/instrumentación , Fotometría/instrumentación , Procesamiento de Señales Asistido por Computador/instrumentación , Aumento de la Imagen/instrumentación , Fotometría/métodos , Fotones , Reproducibilidad de los Resultados , Sensibilidad y EspecificidadRESUMEN
In this paper, a practical continuous-variable quantum key distribution system is developed and it runs in the real-world conditions with 25 MHz clock rate. To reach high-rate, we have employed a homodyne detector with maximal bandwidth to 300 MHz and an optimal high-efficiency error reconciliation algorithm with processing speed up to 25 Mbps. To optimize the stability of the system, several key techniques are developed, which include a novel phase compensation algorithm, a polarization feedback algorithm, and related stability method on the modulators. Practically, our system is tested for more than 12 hours with a final secret key rate of 52 kbps over 50 km transmission distance, which is the highest rate so far in such distance. Our system may pave the road for practical broadband secure quantum communication with continuous variables in the commercial conditions.
RESUMEN
Considering a practical continuous variable quantum key distribution(CVQKD) system, synchronization is of significant importance as it is hardly possible to extract secret keys from unsynchronized strings. In this paper, we proposed a high performance frame synchronization method for CVQKD systems which is capable to operate under low signal-to-noise(SNR) ratios and is compatible with random phase shift induced by quantum channel. A practical implementation of this method with low complexity is presented and its performance is analysed. By adjusting the length of synchronization frame, this method can work well with large range of SNR values which paves the way for longer distance CVQKD.
RESUMEN
In this paper, a hierarchical multi-task structural learning algorithm is developed to support large-scale plant species identification, where a visual tree is constructed for organizing large numbers of plant species in a coarse-to-fine fashion and determining the inter-related learning tasks automatically. For a given parent node on the visual tree, it contains a set of sibling coarse-grained categories of plant species or sibling fine-grained plant species, and a multi-task structural learning algorithm is developed to train their inter-related classifiers jointly for enhancing their discrimination power. The inter-level relationship constraint, e.g., a plant image must first be assigned to a parent node (high-level non-leaf node) correctly if it can further be assigned to the most relevant child node (low-level non-leaf node or leaf node) on the visual tree, is formally defined and leveraged to learn more discriminative tree classifiers over the visual tree. Our experimental results have demonstrated the effectiveness of our hierarchical multi-task structural learning algorithm on training more discriminative tree classifiers for large-scale plant species identification.
Asunto(s)
Algoritmos , Inteligencia Artificial , Plantas/clasificación , ClasificaciónRESUMEN
We report the first continuous-variable quantum key distribution (CVQKD) experiment to enable the creation of 1 Mbps secure key rate over 25 km standard telecom fiber in a coarse wavelength division multiplexers (CWDM) environment. The result is achieved with two major technological advances: the use of a 1 GHz shot-noise-limited homodyne detector and the implementation of a 50 MHz clock system. The excess noise due to noise photons from local oscillator and classical data channels in CWDM is controlled effectively. We note that the experimental verification of high-bit-rate CVQKD in the multiplexing environment is a significant step closer toward large-scale deployment in fiber networks.