Search

Despite their outstanding accuracy, semi-supervised segmentation methods based on deep neural networks can still yield predictions that are considered anatomically impossible by clinicians, for instance, containing holes or disconnected regions. To solve this problem, we present a Context-aware Virtual Adversarial Training (CaVAT) method for generating anatomically plausible segmentation. Unlike approaches focusing solely on accuracy, our method also considers complex topological constraints like connectivity which cannot be easily modeled in a differentiable loss function. We use adversarial training to generate examples violating the constraints, so the network can learn to avoid making such incorrect predictions on new examples, and employ the Reinforce algorithm to handle non-differentiable segmentation constraints. The proposed method offers a generic and efficient way to add any constraint on top of any segmentation network. Experiments on two clinically-relevant datasets show our method to produce segmentations that are both accurate and anatomically-plausible in terms of region connectivity.

Deep co-training has recently been proposed as an effective approach for image segmentation when annotated data is scarce. In this paper, we improve existing approaches for semi-supervised segmentation with a self-paced and self-consistent co-training method. To help distillate information from unlabeled images, we first design a self-paced learning strategy for co-training that lets jointly-trained neural networks focus on easier-to-segment regions first, and then gradually consider harder ones. This is achieved via an end-to-end differentiable loss in the form of a generalized Jensen Shannon Divergence (JSD). Moreover, to encourage predictions from different networks to be both consistent and confident, we enhance this generalized JSD loss with an uncertainty regularizer based on entropy. The robustness of individual models is further improved using a self-ensembling loss that enforces their prediction to be consistent across different training iterations. We demonstrate the potential of our method on three challenging image segmentation problems with different image modalities, using a small fraction of labeled data. Results show clear advantages in terms of performance compared to the standard co-training baselines and recently proposed state-of-the-art approaches for semi-supervised segmentation.

We present the design and experimental demonstration of a compact tunable silicon photonic interleaver with a wide spectral range. The interleaver consists of an interfering loop containing a Fabry–Perot cavity formed by two Sagnac loops. The transmission coefficients of the directional couplers are calculated based on the maximally flat criterion. The impacts of the transmission coefficient variations on the extinction ratio and filter sharpness are investigated. The wavelength dependences of the directional couplers are analyzed to increase the spectral range of the interleaver. The device was fabricated on a silicon-on-insulator platform. The spectral ranges of the measured transmission and reflection spectra are increased to 60 nm. By thermal tuning the waveguide connecting the two Sagnac loops, the central wavelength can be shifted over one free spectral range with a wavelength-tuning efficiency of ∼0.08 nm/mW.

An efficient strategy for weakly-supervised segmentation is to impose constraints or regularization priors on target regions. Recent efforts have focused on incorporating such constraints in the training of convolutional neural networks (CNN), however this has so far been done within a continuous optimization framework. Yet, various segmentation constraints and regularization priors can be modeled and optimized more efficiently in a discrete formulation. This paper proposes a method, based on the alternating direction method of multipliers (ADMM) algorithm, to train a CNN with discrete constraints and regularization priors. This method is applied to the segmentation of medical images with weak annotations, where both size constraints and boundary length regularization are enforced. Experiments on two benchmark datasets for medical image segmentation show our method to provide significant improvements compared to existing approaches in terms of segmentation accuracy, constraint satisfaction and convergence speed.

In this work, a new identification method for the load identification of ball mill based on improved K-means clustering is proposed. Since the external device response signals of ball mill are closely related to its load during the operation, the appropriate parameters are selected. The number of clusters and initial clustering centers are meliorated based on their density and genetic algorithm. The dimensionality of input features are reduced with kernel principal component analysis (KPCA), and the final clustering centers are obtained by iterating in the space with reduced dimensionalities. Tests on datasets have demonstrated the higher accuracy and stability of our improved K-means method compared with other methods. The case-analysis results show that the mill load status can be identified clearly according to the displacement variations of cluster centers calculated by this method. Our research provides a new reference for mill load identification in the viewpoint of multi-source information fusion.

In this paper, we aim to improve the performance of semantic image segmentation in a semi-supervised setting where training is performed with a reduced set of annotated images and additional non-annotated images. We present a method based on an ensemble of deep segmentation models. Models are trained on subsets of the annotated data and use non-annotated images to exchange information with each other, similar to co-training. Diversity across models is enforced with the use of adversarial samples. We demonstrate the potential of our method on three challenging image segmentation problems, and illustrate its ability to share information between simultaneously trained models, while preserving their diversity. Results indicate clear advantages in terms of performance compared to recently proposed semi-supervised methods for segmentation.

We propose and experimentally demonstrate a photonic generation scheme for multiband ultra-wideband (MB-UWB) signal by employing a dual-drive Mach–Zehnder modulator (DDMZM). In this scheme, four electrical binary pulse streams with different bit rates are combined to drive one radio frequency (RF) port of the DDMZM, while the other RF port is modulated by the sum of their 1-b delayed duplicates. By properly adjusting the bit rate and the amplitude of each pulse stream, as well as the bias voltage of the DDMZM, an MB-UWB signal with four sub-bands can be obtained and the transmission capacity can be quadrupled. The feasibility of the proposed scheme is verified by a system demonstration, where the signal modulated by amplitude-phase-shift-key format reaches a high bit rate of 2 Gb/s.

We present our recent progress on microwave signal processing (MSP) using on-chip passive silicon photonic devices, including tunable microwave notch filtering/millimeter-wave (MMW) signal generation based on self-coupled micro-resonators (SCMRs), and tunable radio-frequency (RF) phase shifting implemented by a micro-disk resonator (MDR). These schemes can provide improved flexibility and performances of MSP. The experimental results are in good agreement with theoretical predictions, which validate the effectiveness of the proposed schemes.

We propose and experimentally demonstrate an on-chip all-optical differential-equation solver capable of solving second-order ordinary differential equations (ODEs) characterizing continuous-time linear time-invariant systems. The photonic device is implemented by a self-coupled microresonator on a silicon-on-insulator platform with mutual coupling between the cavity modes. Owing to the mutual mode coupling within the same resonant cavity, the resonance wavelengths induced by different cavity modes are self-aligned, thus avoiding precise wavelength alignment and unequal thermal wavelength drifts as in the case of cascaded resonators. By changing the mutual mode coupling strength, the proposed device can be used to solve second-order ODEs with tunable coefficients. System demonstration using the fabricated device is carried out for 10-Gb/s optical Gaussian and super-Gaussian input pulses. The experimental results are in good agreement with theoretical predictions of the solutions, which verify the feasibility of the fabricated device as a tunable second-order photonic ODE solver.

A compact silicon bandpass filter with high sidelobe suppression is proposed and experimentally demonstrated using an apodized subwavelength grating (SWG) coupler. The device is implemented by placing a SWG waveguide next to a strip waveguide, and apodization is employed with a Gaussian profile to taper the gap between the two waveguides. A high sidelobe suppression ratio of 27 dB can be obtained with a 3-dB bandwidth of 8.8 nm and an insertion loss of 2.5 dB. Owing to the large optical phase mismatch between the two waveguides and the presence of the SWG waveguide, the coupling length of the device is reduced to 100.3 μm. The experimental results validate our proposed apodized-SWG-based contradirectional coupler (contra-DC) as a promising device in suppressing out-of-band components in coarse wavelength division multiplexed (CWDM) optical communication systems.

An effective sensitivity-improving scheme combining symbol-amplitude scaling and peak-amplitude reduction is proposed for IM-DD OFDM-PONs. Sensitivity improvements of 3.8 dB and 4 dB are observed in two ONUs with 40-km and 80-km transmission distances, respectively.

Oligomeric nitrilase was fused with an amphipathic self-assembly peptide 18A at the C-terminus and expressed in Escherichia coli. The fusion enzyme spontaneously assembled into active aggregates with >90% native nitrilase activity. Much higher specific activities than the native nitrilase (Nit) were recorded for the fusion nitrilase (Nit-SEA) at higher temperatures. The enzyme aggregates were purified through cell lysis and centrifugation led to the facile preparation of immobilized particles (Nit-iSEA). Approximately 86% of the initial nitrilase activity was incorporated into the Ca-alginate entrapment beads. The thermostability of the four kinds of different nitrilase variants showed that the Nit-SEA and Nit-iSEA at 45 °C were about 6.7- and 10-fold more stable than the native nitrilase. The nitrile tolerance was also dramatically improved, no apparent substrate inhibition was observed for Nit-iSEA over the range of 30 to 120 mM mandelonitrile. Additionally, the Nit-iSEA could be recycled 20 times with ∼5% loss in activity.

We propose a compact (5μm×40μm) 2×2 electro-optic switch implemented by dual silicon photonic crystal nanobeam (PCN) cavities. A switching power of 2.6 fJ/bit is numerically demonstrated. Extinction ratios of 14.2/13.4 dB are achieved at through/drop ports.

We propose and experimentally demonstrate an on-chip silicon photonic differential-equation solver that can be used to solve second-order constant-coefficient ordinary-differential-equations (ODEs). The feasibility of the device is verified by 5 and 10 Gb/s optical signals.

We propose and experimentally demonstrate a new scheme for orthogonal frequency division multiplexing (OFDM) sidelobe suppression by inserting sidelobe suppression prefix (SSP) ahead of cyclic prefix (CP). The feasibility of the proposal is verified by both simulation and experiment.

A simple and effective scheme of symbol-amplitude scaling is proposed and experimentally demonstrated to significantly improve OFDM transmitter performance. Up to 4.3-dB sensitivity improvement is achieved for OFDM QPSK signal in a LR-PON scenario.

We propose and numerically demonstrate a scheme that enables the direct detection of QAM signals by a single photodetector. 40-Gb/s 16-QAM signal over 22-km fiber transmission is recovered without sacrificing optical spectral efficiency.