Masked autoencoders with generalizable self-distillation for skin lesion segmentation.

In the field of skin lesion image segmentation, accurate identification and partitioning of diseased regions is of vital importance for in-depth analysis of skin cancer. Self-supervised learning, i.e., MAE, has emerged as a potent force in the medical imaging domain, which autonomously learns and extracts latent features from unlabeled data, thereby yielding pre-trained models that greatly assist downstream tasks. To encourage pre-trained models to more comprehensively learn the global structural and local detail information inherent in dermoscopy images, we introduce a Teacher-Student architecture, named TEDMAE, by incorporating a self-distillation mechanism, it learns holistic image feature information to improve the generalizable global knowledge learning of the student MAE model. To make the image features learned by the model suitable for unknown test images, two optimization strategies are, Exterior Conversion Augmentation (EC) utilizes random convolutional kernels and linear interpolation to effectively transform the input image into one with the same shape but altered intensities and textures, while Dynamic Feature Generation (DF) employs a nonlinear attention mechanism for feature merging, enhancing the expressive power of the features, are proposed to enhance the generalizability of global features learned by the teacher model, thereby improving the overall generalization capability of the pre-trained models. Experimental results from the three public skin disease datasets, ISIC2019, ISIC2017, and PH 2 indicate that our proposed TEDMAE method outperforms several similar approaches. Specifically, TEDMAE demonstrated optimal segmentation and generalization performance on the ISIC2017 and PH 2 datasets, with Dice scores reaching 82.1% and 91.2%, respectively. The best Jaccard values were 72.6% and 84.5%, while the optimal HD95% values were 13.0% and 8.9%, respectively.

A Message Passing-Assisted Iterative Noise Cancellation Method for Clipped OTFS-BFDM Systems.

Compared with orthogonal frequency division multiplexing (OFDM) systems, orthogonal time frequency space systems based on bi-orthogonal frequency division multiplexing (OTFS-BFDM) have lower out-of-band emission (OOBE) and better robustness to high-mobility scenarios, but suffer from a higher peak-to-average ratio (PAPR) in large data packets. In this paper, one-iteration clipping and filtering (OCF) is adopted to reduce the PAPR of OTFS-BFDM signals. However, the extra noise introduced by the clipping process, i.e., clipping noise, will distort the desired signal and increase the bit error rate (BER). We propose a message passing (MP)-assisted iterative cancellation (MP-AIC) method to cancel the clipping noise based on the traditional MP decoding at the receiver, which incorporates with the (OCF) at the transmitter to keep the sparsity of the effective channel matrix. The main idea of MP-AIC is to extract the residual signal fed to the MP detector by iteratively constructing reference clipping noise at the receiver. During each iteration, the variance of residual signal and channel noise are taken as input parameters of MP decoding to improve the BER. Moreover, the convergence probability of the modulation alphabet after MP decoding in the current iteration is used as the initial probability of MP decoding in the next iteration to accelerate the convergence rate of MP decoding. Simulation results show that the proposed MP-AIC method significantly improves MP-decoding accuracy while accelerating the BER convergence in the clipped OTFS-BFDM system. In the clipped OTFS-BFDM system with rectangular pulse shaping, the BER of MP-AIC with two iterations can be reduced by 72% more than that without clipping noise cancellation.

A Low Collision and High Throughput Data Collection Mechanism for Large-Scale Super Dense Wireless Sensor Networks.

Super dense wireless sensor networks (WSNs) have become popular with the development of Internet of Things (IoT), Machine-to-Machine (M2M) communications and Vehicular-to-Vehicular (V2V) networks. While highly-dense wireless networks provide efficient and sustainable solutions to collect precise environmental information, a new channel access scheme is needed to solve the channel collision problem caused by the large number of competing nodes accessing the channel simultaneously. In this paper, we propose a space-time random access method based on a directional data transmission strategy, by which collisions in the wireless channel are significantly decreased and channel utility efficiency is greatly enhanced. Simulation results show that our proposed method can decrease the packet loss rate to less than 2 % in large scale WSNs and in comparison with other channel access schemes for WSNs, the average network throughput can be doubled.

An Adaptive Channel Access Method for Dynamic Super Dense Wireless Sensor Networks.

Super dense and distributed wireless sensor networks have become very popular with the development of small cell technology, Internet of Things (IoT), Machine-to-Machine (M2M) communications, Vehicular-to-Vehicular (V2V) communications and public safety networks. While densely deployed wireless networks provide one of the most important and sustainable solutions to improve the accuracy of sensing and spectral efficiency, a new channel access scheme needs to be designed to solve the channel congestion problem introduced by the high dynamics of competing nodes accessing the channel simultaneously. In this paper, we firstly analyzed the channel contention problem using a novel normalized channel contention analysis model which provides information on how to tune the contention window according to the state of channel contention. We then proposed an adaptive channel contention window tuning algorithm in which the contention window tuning rate is set dynamically based on the estimated channel contention level. Simulation results show that our proposed adaptive channel access algorithm based on fast contention window tuning can achieve more than 95 % of the theoretical optimal throughput and 0 . 97 of fairness index especially in dynamic and dense networks.

Homotopic non-local regularized reconstruction from sparse positron emission tomography measurements.

BACKGROUND: Positron emission tomography scanners collect measurements of a patient's in vivo radiotracer distribution. The system detects pairs of gamma rays emitted indirectly by a positron-emitting radionuclide (tracer), which is introduced into the body on a biologically active molecule, and the tomograms must be reconstructed from projections. The reconstruction of tomograms from the acquired PET data is an inverse problem that requires regularization. The use of tightly packed discrete detector rings, although improves signal-to-noise ratio, are often associated with high costs of positron emission tomography systems. Thus a sparse reconstruction, which would be capable of overcoming the noise effect while allowing for a reduced number of detectors, would have a great deal to offer. METHODS: In this study, we introduce and investigate the potential of a homotopic non-local regularization reconstruction framework for effectively reconstructing positron emission tomograms from such sparse measurements. RESULTS: Results obtained using the proposed approach are compared with traditional filtered back-projection as well as expectation maximization reconstruction with total variation regularization. CONCLUSIONS: A new reconstruction method was developed for the purpose of improving the quality of positron emission tomography reconstruction from sparse measurements. We illustrate that promising reconstruction performance can be achieved for the proposed approach even at low sampling fractions, which allows for the use of significantly fewer detectors and have the potential to reduce scanner costs.

Noise-compensated homotopic non-local regularized reconstruction for rapid retinal optical coherence tomography image acquisitions.

BACKGROUND: Optical coherence tomography (OCT) is a minimally invasive imaging technique, which utilizes the spatial and temporal coherence properties of optical waves backscattered from biological material. Recent advances in tunable lasers and infrared camera technologies have enabled an increase in the OCT imaging speed by a factor of more than 100, which is important for retinal imaging where we wish to study fast physiological processes in the biological tissue. However, the high scanning rate causes proportional decrease of the detector exposure time, resulting in a reduction of the system signal-to-noise ratio (SNR). One approach to improving the image quality of OCT tomograms acquired at high speed is to compensate for the noise component in the images without compromising the sharpness of the image details. METHODS: In this study, we propose a novel reconstruction method for rapid OCT image acquisitions, based on a noise-compensated homotopic modified James-Stein non-local regularized optimization strategy. The performance of the algorithm was tested on a series of high resolution OCT images of the human retina acquired at different imaging rates. RESULTS: Quantitative analysis was used to evaluate the performance of the algorithm using two state-of-art denoising strategies. Results demonstrate significant SNR improvements when using our proposed approach when compared to other approaches. CONCLUSIONS: A new reconstruction method based on a noise-compensated homotopic modified James-Stein non-local regularized optimization strategy was developed for the purpose of improving the quality of rapid OCT image acquisitions. Preliminary results show the proposed method shows considerable promise as a tool to improve the visualization and analysis of biological material using OCT.

Homotopic, non-local sparse reconstruction of optical coherence tomography imagery.

The resolution in optical coherence tomography imaging is an important parameter which determines the size of the smallest features that can be visualized. Sparse sampling approaches have shown considerable promise in producing high resolution OCT images with fewer camera pixels, reducing both the cost and the complexity of an imaging system. In this paper, we propose a non-local approach to the reconstruction of high resolution OCT images from sparsely sampled measurements. An iterative strategy is introduced for minimizing a homotopic, non-local regularized functional in the spatial domain, subject to data fidelity constraints in the k-space domain. The novel algorithm was tested on human retinal, corneal, and limbus images, acquired in-vivo, demonstrating the effectiveness of the proposed approach in generating high resolution reconstructions from a limited number of camera pixels.