Ensuring fairness in medical image segmentation is critical due to biases in imbalanced clinical data acquisition caused by demographic attributes (e.g., age, sex, race) and clinical factors (e.g., disease severity). To address these challenges, we introduce Distribution-aware Mixture of Experts (dMoE), inspired by optimal control theory. We provide a comprehensive analysis of its underlying mechanisms and clarify dMoE's role in adapting to heterogeneous distributions in medical image segmentation. Furthermore, we integrate dMoE into multiple network architectures, demonstrating its broad applicability across diverse medical image analysis tasks. By incorporating demographic and clinical factors, dMoE achieves state-of-the-art performance on two 2D benchmark datasets and a 3D in-house dataset. Our results highlight the effectiveness of dMoE in mitigating biases from imbalanced distributions, offering a promising approach to bridging control theory and medical image segmentation within fairness learning paradigms. The source code is available at https://github.com/tvseg/dMoE.

Recent advances in AI foundation models have significant potential for lightening the clinical workload by mimicking the comprehensive and multi-faceted approaches used by medical professionals. In the field of radiation oncology, the integration of multiple modalities holds great importance, so the opportunity of foundational model is abundant. Inspired by this, here we present RO-LMM, a multi-purpose, comprehensive large multimodal model (LMM) tailored for the field of radiation oncology. This model effectively manages a series of tasks within the clinical workflow, including clinical context summarization, radiotherapy strategy suggestion, and plan-guided target volume segmentation by leveraging the capabilities of LMM. In particular, to perform consecutive clinical tasks without error accumulation, we present a novel Consistency Embedding Fine-Tuning (CEFTune) technique, which boosts LMM's robustness to noisy inputs while preserving the consistency of handling clean inputs. We further extend this concept to LMM-driven segmentation framework, leading to a novel Consistency Embedding Segmentation (CESEG) techniques. Experimental results including multi-center validation confirm that our RO-LMM with CEFTune and CESEG results in promising performance for multiple clinical tasks with generalization capabilities.

Automatic Speech Recognition (ASR) is a technology that converts spoken words into text, facilitating interaction between humans and machines. One of the most common applications of ASR is Speech-To-Text (STT) technology, which simplifies user workflows by transcribing spoken words into text. In the medical field, STT has the potential to significantly reduce the workload of clinicians who rely on typists to transcribe their voice recordings. However, developing an STT model for the medical domain is challenging due to the lack of sufficient speech and text datasets. To address this issue, we propose a medical-domain text correction method that modifies the output text of a general STT system using the Vision Language Pre-training (VLP) method. VLP combines textual and visual information to correct text based on image knowledge. Our extensive experiments demonstrate that the proposed method offers quantitatively and clinically significant improvements in STT performance in the medical field. We further show that multi-modal understanding of image and text information outperforms single-modal understanding using only text information.

Despite promising advancements in deep learning in medical domains, challenges still remain owing to data scarcity, compounded by privacy concerns and data ownership disputes. Recent explorations of distributed-learning paradigms, particularly federated learning, have aimed to mitigate these challenges. However, these approaches are often encumbered by substantial communication and computational overhead, and potential vulnerabilities in privacy safeguards. Therefore, we propose a self-supervised masked sampling distillation technique called MS-DINO, tailored to the vision transformer architecture. This approach removes the need for incessant communication and strengthens privacy using a modified encryption mechanism inherent to the vision transformer while minimizing the computational burden on client-side devices. Rigorous evaluations across various tasks confirmed that our method outperforms existing self-supervised distributed learning strategies and fine-tuned baselines.

The widespread application of artificial intelligence in health research is currently hampered by limitations in data availability. Distributed learning methods such as federated learning (FL) and split learning (SL) are introduced to solve this problem as well as data management and ownership issues with their different strengths and weaknesses. The recent proposal of federated split task-agnostic (F eSTA) learning tries to reconcile the distinct merits of FL and SL by enabling the multi-task collaboration between participants through Vision Transformer (ViT) architecture, but they suffer from higher communication overhead. To address this, here we present a multi-task distributed learning using ViT with random patch permutation, dubbed p -F eSTA. Instead of using a CNN-based head as in F eSTA, p -F eSTA adopts a simple patch embedder with random permutation, improving the multi-task learning performance without sacrificing privacy. Experimental results confirm that the proposed method significantly enhances the benefit of multi-task collaboration, communication efficiency, and privacy preservation, shedding light on practical multi-task distributed learning in the field of medical imaging.

Federated learning, which shares the weights of the neural network across clients, is gaining attention in the healthcare sector as it enables training on a large corpus of decentralized data while maintaining data privacy. For example, this enables neural network training for COVID-19 diagnosis on chest X-ray (CXR) images without collecting patient CXR data across multiple hospitals. Unfortunately, the exchange of the weights quickly consumes the network bandwidth if highly expressive network architecture is employed. So-called split learning partially solves this problem by dividing a neural network into a client and a server part, so that the client part of the network takes up less extensive computation resources and bandwidth. However, it is not clear how to find the optimal split without sacrificing the overall network performance. To amalgamate these methods and thereby maximize their distinct strengths, here we show that the Vision Transformer, a recently developed deep learning architecture with straightforward decomposable configuration, is ideally suitable for split learning without sacrificing performance. Even under the non-independent and identically distributed data distribution which emulates a real collaboration between hospitals using CXR datasets from multiple sources, the proposed framework was able to attain performance comparable to data-centralized training. In addition, the proposed framework along with heterogeneous multi-task clients also improves individual task performances including the diagnosis of COVID-19, eliminating the need for sharing large weights with innumerable parameters. Our results affirm the suitability of Transformer for collaborative learning in medical imaging and pave the way forward for future real-world implementations.

Under the global pandemic of COVID-19, the use of artificial intelligence to analyze chest X-ray (CXR) image for COVID-19 diagnosis and patient triage is becoming important. Unfortunately, due to the emergent nature of the COVID-19 pandemic, a systematic collection of CXR data set for deep neural network training is difficult. To address this problem, here we propose a patch-based convolutional neural network approach with a relatively small number of trainable parameters for COVID-19 diagnosis. The proposed method is inspired by our statistical analysis of the potential imaging biomarkers of the CXR radiographs. Experimental results show that our method achieves state-of-the-art performance and provides clinically interpretable saliency maps, which are useful for COVID-19 diagnosis and patient triage.

Although there are many commercial products and services advertising their capability of indoor navigation using smartphone, still the service coverage is highly restricted to the specific area due to tremendous effort on pre-survey and maintenance on positioning resources such as radio map of Wi-Fi signals. Most of indoor positioning is dependent on the installed facilities such as Wi-Fi APs and BLE beacons in service area. Without pre-surveyed information such as fingerprint database, one can not localize himself at the first visiting site.

Ultrawideband (UWB) based localization has the potential to be used in a variety of applications due to its high accuracy. For robust and high performance in real environment, the most challenging issue is the detection and mitigation of noise from non-line-of-sight (NLOS) signals. Current researches use channel state information, particle or Kalman filter, and statistics based approaches for the NLOS noise detection and mitigation. These solutions show high accuracy in some applications; however, they need additional hardware and work in static environment only. We propose an NLOS mitigation algorithm that does not need any additional hardware andworks in dynamic environment with mobile obstacles. The main idea of the algorithm are to estimate the NLOS bias by repetitively comparing the intersection of the hyperbolas with intersections of circles. The proposed approach is tested for Decawave UWB testbed. Experimental results show that the proposed scheme works well in different dynamic scenarios as compared to the localization scheme without NLOS noise mitigation.

In this study, the authors propose a low‐complexity symbol‐level combining (SLC) scheme for Chase combining‐based hybrid automatic repeat request in multiple‐input multiple‐output systems with a linear receiver. In the proposed scheme, a subset of all row vectors in the channel matrix is selected and used for combining and detection of retransmitted symbols, where the Sherman–Morrison–Woodbury lemma is applied for the utilisation of each selected row vector. Therefore, compared with the existing SLC schemes using the entire channel matrix, the proposed scheme can have an advantage in computational complexity. Furthermore, they develop a row vector selection criterion for the proposed SLC scheme to calculate the amount of the signal‐to‐interference‐noise ratio improvement based on the squared norm of each row vector with a significantly low complexity. Simulation results verify that compared with the existing SLC schemes, the proposed SLC scheme achieves similar or better error performance, while its computational complexity is lower or in the worst‐case similar.

The performance of symbol-level combining (SLC) and bit-level combining (BLC) in multiple-input multiple-output (MIMO) multiple automatic repeat request (ARQ) systems is investigated. Considering the zero-forcing (ZF) detection under the assumption of perfect packet elimination for SLC, the two performance characteristics of Chase combining (CC) with SLC and BLC in MIMO multiple ARQ (MMARQ) systems are analyzed. First, the performance of a packet with the highest hybrid ARQ (HARQ) round in CC-SLC-ZF is not affected by the HARQ rounds of the other packets simultaneously sent. Second, the performance gain of CC-SLC-ZF over CC-BLC-ZF for a packet can be improved when the packets with higher HARQ rounds are simultaneously sent. The latter indicates that CC-SLC can provide an improved error performance for the packets transmitted only once when there is at least one retransmitted packet simultaneously sent. Therefore, even though incremental redundancy (IR) provides a significantly larger coding gain than CC, CC-SLC can provide a better throughput than IR-BLC as the average block error rate of a retransmitted packet approaches zero. Simulation results verify that the analyses remain valid regardless of the detection scheme and the throughput of CC-SLC at the high SNR region can be better than IR-BLC in MMARQ systems.

In this paper, we consider a full-duplex massive MIMO system which includes a full-duplex BS with a large number of antennas and half-duplex uplink (UL) and downlink (DL) users with a single antenna. Due to the self-interference at the BS, the UL spectral efficiency is significantly decreased. In order to improve the UL spectral efficiency, we focus on the utilization schemes of the spatial resources which can be used to enhance the UL performance. If the BS uses the spatial resources as receive antennas, the BS can obtain an additional receive diversity gain for the UL desired signal. On the other hand, if the BS uses the spatial resources as transmit antennas, the BS can suppress a part of the self-interference by using the extended transmit beamformer including a part of the self-interference channel. Our analysis and numerical results show that there are some tradeoffs between achieving receive diversity gain for the UL desired signal and suppressing the self-interference. According to the system parameters, the utilization schemes of the spatial resources would be determined for maximizing the UL spectral efficiency.

This paper analyzes the ratio of the transmit antennas and receive antennas in multi-user multiple-input multipleoutput with a full-duplex and large array base station (BS) and half-duplex users (MU-MIMO FLB-HU) systems. We consider the BS exploits zero-forcing beamformer and zero-forcing receiver. We derive the deterministic approximation of the downlink and uplink sum-rates considering inter-user interference and self-interference, respectively. Based on the analyzed results, we formulate an optimization problem in terms of the number of transmit and receive antennas to maximize the sum of downlink and uplink sum-rates subject to the number of total antennas at the BS. From the optimization problem, the optimal antenna ratio between the number of transmit and receive antennas can be obtained. We analyze that the optimal antenna ratio converges to the ratio between the number of downlink users (Kd) and uplink users (Ku) as the number of total antennas goes to infinity. Simulation results show that the optimal antenna ratio enhances the sum-rate performance compared to the same number of transmit antennas and receive antennas in the MU-MIMO FLBHU system. In particular, in the MU-MIMO FLB-HU system with Kd= 10 and Ku= 5, the optimal antenna ratio can achieve about 5∼10bps/Hz performance gain compared to the same number of transmit antennas and receive antennas.

The goal of this paper is to develop a hybrid automatic repeat request (HARQ) transmission state control algorithm for multiple-input multiple- output (MIMO) systems with interference cancellation (IC) detection to improve throughput. The HARQ transmission state is defined as the distribution of the initial packets and retransmission packets transmitted during a transmission time slot. In such a system, the retransmission packets normally have a lower error probability than initial packets. However, in terms of throughput, successful decoding of an initial packet is more rewarding than a retransmission packet. The proposed algorithm generates the transmission state in which initial packets and retransmission packets are sent together. The outcome is that it achieves a lower error probability for initial packets by exploiting the IC process and a significantly higher throughput than the conventional HARQ system, which is verified by simulation results.

A packet transmission strategy for multiple‐input multiple‐output systems with interference cancellation detection and hybrid automatic repeat request (HARQ) processes is proposed. The error probability and throughput of such a system depend on the choice of the set of packets to be transmitted in each packet transmission time interval, but there are no existing implementable strategies to optimise such choice of packets. In this study, the authors first define a concept of the effective interference level (EIL) and establish the relationship between the EIL and the effective signal‐to‐interference‐plus‐noise ratio. Then they show that choosing the set of packets that minimises the EIL successively from the lowest to the highest HARQ round leads to a lower packet error and higher system throughput than conventional HARQ, which is verified by simulation. Also, the proposed EIL‐based scheme uses only the acknowledgement feedback messages like a conventional HARQ scheme, because the number of HARQ rounds of each packet is the only required information to calculate the EIL. Simulation results show that the proposed scheme outperforms the conventional scheme in terms of throughput with the signal‐to‐noise ratio gain of about 4.2 dB at maximum.

In this paper, we proposed Advanced Heuristic Drift Elimination (AHDE) which can remove azimuth drift error in indoor environments. In Pedestrian Dead Reckoning (PDR) system, azimuth error is one of the main factors that cause estimated position error. In order to reduce azimuth error, several methods are used. Heuristic Drift Elimination (HDE) algorithm proposed by Johann Borenstein shows great strength in indoor environments. HDE assumes that generally walls and corridors are straight and either parallel or orthogonal to each other in man-made building. They called the typical directions of walls and corridors as the dominant directions. HDE is corrected if the computed azimuth angle matches the closest dominant direction. HDE also has limitation when the pedestrian walks in various directions because HDE can cause a new azimuth error by matching the closed dominant direction. To overcome these limitations, we propose AHDE which is based on INS-EKF-ZUPT (IEZ) by using foot-mounted IMU. The algorithm consists with the following two steps. First, it determines whether a pedestrian is walking straight forward or not. If a pedestrian is not walking straight forward, the algorithm estimates the biases of accelerometers and gyroscopes by Zero velocity UPdaTe (ZUPT) method. However if the pedestrian is walking straight forward, the algorithm determines whether the pedestrian is walking along the dominant direction or not. When it is determined that pedestrian is walking along the dominant direction, the algorithm corrects the computed azimuth angle to the closest dominant direction. When it is determined that the pedestrian is not walking along the dominant direction but walking straight with no change in azimuth, AHDE applies a correction to the gyro output which contains the bias error. Experimental results show that the accuracy of AHDE is improved compared to HDE and the algorithm is a powerful method which can reduce the azimuth error in complex motion.

Wi-Fi-based localization is one of the most promising technologies for indoor location-based services. However, it is still a difficult task to construct a positioning database to provide high accuracy and vast coverage, especially in probabilistic-based algorithms such as fingerprint-based localization. On the other hand, positioning methods that use the location database of the positioning infrastructure, such as the weighted centroid method, can support terminal-based localization due to lower computational complexity and small database size. Terminal-based positioning has the advantages not only of protecting personal location information, but also of maintaining the network topology among the terminals. The only weak point for positioning database-based methods is that the accurate location of the infrastructure should be known in advance. In this paper, we propose a novel algorithm to estimate the location of infrastructure, especially for Wi-Fi access points. We developed and employed a smartphone application to collect Wi-Fi signals by just walking around the test area. Then, the locations of Wi-Fi access points are estimated efficiently and accurately by selection of the maximum likelihood position that has the most similar path-loss model corresponding to the signal acquisition points. The estimated locations are processed as infrastructure database to support terminal-based positioning. The simulation and experiment result validate the feasibility of the proposed algorithm. The estimated location of access points is within 10 m accuracy in most cases and also the terminal-based positioning results achieve competitive performance comparing with other positioning methods.

We propose and demonstrate an adaptive wireless mesh network architecture for public surveillance and monitoring. The proposed architecture is based on IEEE 802.11s WLAN mesh network, which is enhanced to provide adaptation depending on the environment of the network. We first introduce the implementation architecture of the adaptive mesh, and then describe the demonstration of the network on the wireless mesh camera sensor testbed.

In this paper, an extended detection scheme is proposed for multiple-input multiple-output (MIMO) systems which employ partial incremental redundancy (IR) based hybrid automatic repeat request (HARQ). Based on the extended MIMO system model that interprets multiple retransmissions as a single transmission, the extended detection scheme estimates the entire transmitted symbols utilizing all information obtained up to the current retransmission. Compared with the conventional zero-forcing (ZF) detection scheme, the extended ZF detection scheme improves the post-processing signal-to-noise ratio (PSNR) of the newly transmitted symbols in every retransmission, as well as the repeatedly transmitted symbols. Simulation results verify that the extended detection scheme outperforms the conventional detection scheme. Simulation results also show that the extended minimum mean-square-error (MMSE) detection scheme can achieve a better error performance than the conventional maximum-likelihood (ML) detection scheme, and even the extended ZF detection scheme shows a comparable error performance to the conventional ML detection scheme.

In video surveillance and monitoring systems based on wireless mesh networks, large volumes of multimedia data with different priorities and time demands are generated by video cameras and wireless sensors. Hence, one of the key issues is how to alleviate network congestion for highly reliable transfer of multimedia data towards gateways. This paper proposes a solution for utilizing multiple gateways to divert data streams and alleviate traffic congestion near a specific gateway. With a new routing protocol, named “Multi-Gateway Routing with Congestion Avoidance (MGR-CA)”, network congestion can be predicted in a distributed manner and amounts of data traffic transmitted to the congested path are redirected to an alternative gateway based on the severity of congestion. Preliminary testbed experiments are made to show the performance of our scheme.

In this work, an Adaptive Channel Hopping (ACH) mechanism for sensor networks is proposed to avoid interference from other sources and narrow-band jamming. Under unfavorable channel conditions, ACH lets sensors switch to a new operating channel. ACH reduces the channel scanning and selection latency by ordering available channels using link quality indicator measurements and weights. The proposed ACH scheme is evaluated through simulations and a hardware implementation, which suggest low latency and high channel selection quality even in very adverse conditions.

Directional antennas offer many potential advantages for wireless networks such as increased network capacity, extended transmission range and reduced energy consumption. Exploiting these advantages, however, requires new protocols and mechanisms at various communication layers to intelligently control the directional antenna system. With directional antennas, many trivial mechanisms, such as neighbor discovery, become more challenging since communicating parties must agree on where and when to point their directional beams to enable communication. In this paper, we propose a fully directional neighbor discovery protocol called Sectored-Antenna Neighbor Discovery (SAND) protocol. SAND is designed for sectored-antennas, a low-cost and simple realization of directional antennas, that utilize multiple limited beamwidth antennas. Unlike many proposed directional neighbor discovery protocols, SAND depends neither on omnidirectional antennas nor on time synchronization. In addition, SAND performs neighbor discovery in a serialized fashion allowing individual nodes to discover all potential neighbors within a predetermined time. Moreover, SAND guarantees the discovery of the best sector combination on both communication ends allowing more robust and higher reliability links. Finally, SAND gathers the neighborhood information in a centralized location, if needed, to be used by centralized networking protocols. The effectiveness of SAND has been assessed via simulation studies and real hardware implementation.

In numerous sensor network applications, a group of nodes surrounding an event (forming an active region) must collaborate to collect locally generated data, reliably and efficiently. Active regions are often characterized by intense data transfer, packet collisions and congestion, resulting in significant network performance loss. Previous studies dealt these scenarios either at the application or network layers, where potential performance gains at the MAC layer were not considered. In this paper, a cross-layer communication protocol Event-Centric MAC (EC-MAC) is introduced in which the channel access time is divided into two intervals. During slot reservation interval, all nodes within an active region collaborate to construct a Transmit/Receive schedule on-the-fly. Once the schedule is in place, data forwarding is performed during the data transmission interval towards a designated root node. By allowing contention-free channel access, EC-MAC outperforms fully awake and LPL-based MAC protocols in terms of both energy efficiency and data throughput.

Wireless sensor networks have been used to gather data and information in many diverse application settings. The capacity of such networks remains a fundamental obstacle toward the adaptation of sensor network systems for advanced applications that require higher data rates and throughput. In this paper, we explore potential benefits of integrating directional antennas into wireless sensor networks. While the usage of directional antennas has been investigated in the past for ad hoc networks, their usage in sensor networks bring both opportunities as well as challenges. In this paper, Sectored-Antenna Medium Access Control (SAMAC), an integrated cross-layer protocol that provides the communication mechanisms for sensor network to fully utilize sectored antennas, is introduced. Simulation studies show that SAMAC delivers high energy efficiency and predictable delay performance with graceful degradation in performance with increased load.

In patch-and-stitch localization algorithms, a flip error refers to the kind of error in which a patch is stitched to the map as being wrongly reflected. In this paper, we present an anchor-free localization algorithm which tries to detect and prevent flip errors. The flip error prevention is achieved by two filtering mechanisms: the flip ambiguity test and the flip conflict detection. Based on two techniques, we devised an anchor-free localization algorithm and evaluated the performance of the proposed algorithm though simulations. The results show that our algorithm achieves significant performance improvement over the existing algorithms.

In a wireless sensor network application for tracking multiple mobile targets, large amounts of sensing data can be generated by a number of sensors. These data must be controlled with efficient data aggregation techniques to reduce data transmission to the sink node. Several clustering methods were used previously to aggregate the large amounts of data produced from sensors in target tracking applications. However, such clustering based data aggregation algorithms show effectiveness only in restricted type of sensing scenarios, while posing great problems when trying to adapt to various environment changes. To alleviate the problems of existing clustering algorithms, we propose a hybrid clustering based data aggregation scheme. The proposed scheme can adaptively choose a suitable clustering technique depending on the status of the network, increasing the data aggregation efficiency as well as energy consumption and successful data transmission ratio. Performance evaluation via simulation has been made to show the effectiveness of the proposed scheme.

A "patch-and-stitchrdquo localization algorithm divides the network into small overlapping subregions. Typically, each subregion consists of a node and all or some of its neighbors. For each subregion, the algorithm builds a local map, called a patch, which is actually an embedding of the nodes it spans in a relative coordinate system. Finally, the algorithm stitches those patches to form a single global map. In a patch-and-stitch algorithm, the stitching order makes an influence on both the performance and the complexity of the algorithm. In this paper, we present a formal framework to deal with stitching orders in patch-and-stitch localization algorithms. In our framework, the stitching order is determined by a stitching scheme and the stitching scheme consists of a stitching policy and a potential function. The potential function is to predict how well a patch will be stitched if patches are stitched according to a given partial order. The stitching policy is a mechanism that determines the stitching order based on the predictions by the potential function. We present various stitching schemes and evaluate them through simulations. In addition, we apply the patch-and-stitch strategy into the anchor-based localization and propose a clustering-based localization algorithm. A potential function is used to partition the network into clusters each of which is centered at an anchor node. For each cluster, a cluster map is constructed via the anchor-free localization algorithm. Then, those cluster maps are combined to form a single global map. We propose a stitching technique for combining those cluster maps and analyze the performance of the algorithm by simulations.

Data aggregation scheme has been widely used to increase the quality of the target information and to reduce the packet transmissions in the sensor network. In this paper, we provide in-depth study results of various data aggregation schemes through the mathematical analysis and simulations. The results show that the performance of aggregation schemes depends on location of targets, node density, network topology, sensing range, hop counts from sources to the sink. When the hop counts from sources to the sink is small and the sensing range of each node is short, tree-based aggregation scheme achieves better performance than other schemes. As the hop counts from sources to the sink and the sensing range of each node increase, dynamic-cluster aggregation achieves significant performance improvement over the other aggregation schemes