We investigate the effect of explicitly enforcing the Lipschitz continuity of neural networks with respect to their inputs. To this end, we provide a simple technique for computing an upper bound to the Lipschitz constant—for multiple p-norms—of a feed forward neural network composed of commonly used layer types. Our technique is then used to formulate training a neural network with a bounded Lipschitz constant as a constrained optimisation problem that can be solved using projected stochastic gradient methods. Our evaluation study shows that the performance of the resulting models exceeds that of models trained with other common regularisers. We also provide evidence that the hyperparameters are intuitive to tune, demonstrate how the choice of norm for computing the Lipschitz constant impacts the resulting model, and show that the performance gains provided by our method are particularly noticeable when only a small amount of training data is available.

Time of flight (ToF) range cameras illuminate the scene with an amplitude-modulated continuous wave light source and measure the returning modulation envelopes: phase and amplitude. The phase change of the modulation envelope encodes the distance travelled. This technology suffers from measurement errors caused by multiple propagation paths from the light source to the receiving pixel. The multiple paths can be represented as the summation of a direct return, which is the return from the shortest path length, and a global return, which includes all other returns. We develop the use of a sinusoidal pattern from which a closed form solution for the direct and global returns can be computed in nine frames with the constraint that the global return is a spatially lower frequency than the illuminated pattern. In a demonstration on a scene constructed to have strong multipath interference, we find the direct return is not significantly different from the ground truth in $33/136pixels$ tested; where for the full-field measurement, it is significantly different for every pixel tested. The variance in the estimated direct phase and amplitude increases by a factor of eight compared with the standard time of flight range camera technique.

Microwave medical imaging systems have shown a competitive advantage in stroke detection due to their cost-effectiveness, non-ionization, and portability. However, these systems often rely on time-consuming image reconstruction techniques, which are disadvantageous for timely stroke treatment. In this paper, a novel microwave medical sensing (MMS) method is proposed for fast stroke classification and localization, which utilizes space division of the region under examination (i.e., the head). The space division is enabled by the generalized scattering matrix theory and incorporates the brain anatomy. Then a novel decision tree learning method is proposed, which facilitates efficient stroke feature identification for classification. The spatial information acquired from the decision tree also results in rapid stroke localization. In order to verify the proposed method, we investigate the feasibility of classifying brain strokes between an intracranial haemorrhage stroke and an ischemic stroke with a wearable MMS system. Both numerical and experimental results are obtained. Compared to the traditional method, the classification rates for simulation and experimental results are improved by 14.1% and 19.2%, respectively. Furthermore, by utilizing the a priori information, the localization time is reduced by 21.1%. Finally, the localization accuracy is higher than 0.90 in both simulation and experimental studies. The classification accuracy and localization efficiency are shown to be greatly improved compared to the traditional method, which has great significance for wearable devices. This study proposes an efficient space-division-based detection method to localize the brain stroke without imaging.

This letter proposes a general formula to determine the amount of compound required to produce water-oil-based phantoms that can emulate the electrical properties of a specific body tissue at microwave frequencies. The existing process to produce such a phantom usually involves mixing together a number of different materials, where the proportion of each compound needs to be known a priori to reproduce the dielectric properties of a specific tissue. In this letter, we first investigate the dielectric variation of the mixture by changing the proportion of each compound over a wide frequency band. By utilizing Lichtenecker's logarithmic mixture formula, the gradient descent iterative optimization method is applied to train the relations obtained, and a formula that determines the amount of each compound required to produce a phantom given a specific tissue type is derived. The formula is verified by constructing representative tissues of the human head, with the durability and fidelity of the produced phantom also analyzed.

Different natural biological processes are possible because of the collaboration among simple living cells. Similarly, computerized systems, such as multiagent systems (MASs), rely on multiple interacting agents with simplified and reduced capability, to collectively solve difficult problems that are impossible for individual agents to solve on their own. This work highlights an autonomous tumor sensitization strategy in complex human vasculature, where target detection is achieved through the swarm coordination mechanism, with no prior knowledge of tumor location. We propose that small-scale biocompatible organisms, such as nanoparticles, can perform deterministic tasks following the simple principles of aggregation and migration. We aim to show through computational experiments that tumor-triggered biophysical gradients can be leveraged by nanoparticles to collectively move toward the potential tumor hypoxic regions. Although individual nanoparticles have no target-directed locomotion ability due to limited communication and computation capability, we demonstrate that once passive collaboration is achieved, they can successfully avoid obstacles and detect the tumor. Numerical experiments demonstrate that the overall targeting efficiency could improve considerably from 10% to 90% through passive collaboration among nanoparticles. Furthermore, with the introduction of noisy search space and mobile obstacles, the targeting performance would reduce by 25%. Such self-regulating particles can be used as homing agents for target amplification, and hence can assist in early cancer detection through contrast-enhanced medical imaging. We believe that our work will motivate self-dependent and noncentralized approach for magnification of tumor location.

This paper proposes a novel contrast-enhanced microwave cancer detection (MCD) system to accurately detect the location of a tumor targeted by nano-scale contrast agents. This system adopts the angle-of-arrival (AoA) positioning approach that is conventionally applied to the region where plane waves dominate the propagation. Hence, to ensure the effectiveness of the AoA approach when it is working in close proximity to the human body, a new algorithm is proposed to transfer the antenna’s radiation pattern from the far field to the detection region where spherical waves can still be observed. This AoA-based MCD system requires fewer antennas than other radar-based positioning approaches, thereby creating a low-profile hardware architecture. In addition, during the detection process, a differential technique is incorporated in order to track the signal change due to the administration of contrast agents, which can significantly suppress the noise inside the biological medium. The AoA-based differential MCD system is numerically studied on an anatomically realistic breast phantom subject to a variety of signal-to-noise-ratios (SNRs). The results show that the proposed system can successfully locate the tumor with an average resolution of 0.6 mm. When the SNR in the biological medium is 10 dB, the average positioning error is less than 1.5 mm with sensitivity maintained above 60%, which surpasses the performance of other similar systems. The system is also experimentally validated with a physical breast phantom and stepper-motor-driven rotating antennas, where the reconstructed image shows a minor deviation of 3.61 mm with that of the simulated one.