This PDF file contains the Front Matter for SPIE Proceedings Volume 12655, including the Title Page, Copyright information, Table of Contents and Conference Committee list.

Modern deep learning-based cell segmentation algorithms with high computational power have enabled automated and high-throughput segmentation of bacteria. Previous studies, relying on either manual or automated cell segmentation approaches, have proved that a clonal bacteria population cultured in a regular growth medium in a homogenous microenvironment exhibits heterogeneity. When antibiotic treatment has been applied, heterogeneity of the bacteria population may increase depending on the working mechanisms of an administrated antibiotic. Therefore, important features of rare cells, such as asymmetric division of antibiotic persister cells or cells with metastable phenotypes, might be masked by heterogeneity of a population, particularly when a limited number of the cells was analyzed. Therefore, automated image segmentation and analysis approaches have significantly impact on accurate, rapid, and reliable feature identification, particularly for extracting quantitative high-resolution data at high throughput. Here, we implemented U-Net algorithm for segmentation of Escherichia coli cells in the absence and presence of ciprofloxacin. The accuracy values were 0.9912 and 0.9869 for the control and ciprofloxacin-treated cell populations, respectively. Next, we developed an algorithm using Phyton and the OpenCV library to extract the cell number, cellular area, and solidity features of the cells. We believe that our preliminary data might contribute to development of automated, reliable, accurate, and bacteria or antibiotic specific image segmentation tools.

In the past decades, automated high-content microscopy demonstrated its ability to deliver large quantities of image-based data powering the versatility of phenotypic drug screening and systems biology applications. However, as the sizes of image-based datasets grew, it became infeasible for humans to control, avoid and overcome the presence of imaging and sample preparation artefacts in the images. While novel techniques like machine learning and deep learning may address these shortcomings through generative image inpainting, when applied to sensitive research data this may come at the cost of undesired image manipulation. Undesired manipulation may be caused by phenomena such as neural hallucinations, to which some artificial neural networks are prone. To address this, here we evaluate the state-of-the-art inpainting methods for image restoration in a high-content fluorescence microscopy dataset of cultured cells with labelled nuclei. We show that architectures like DeepFill V2 and Edge Connect can faithfully restore microscopy images upon fine-tuning with relatively little data. Our results demonstrate that the area of the region to be restored is of higher importance than shape. Furthermore, to control for the quality of restoration, we propose a novel phenotype-preserving metric design strategy. In this strategy, the size and count of the restored biological phenotypes like cell nuclei are quantified to penalise undesirable manipulation. We argue that the design principles of our approach may also generalise to other applications.

Measuring Hemoglobin (Hb) levels is required for the assessment of different health conditions, such as anemia, a condition where there are insufficient healthy red blood cells to carry enough oxygen to the body’s tissues. Measuring Hb levels requires the extraction of a blood sample, which is then sent to a laboratory for analysis. This is an invasive procedure that may add challenges to the continuous monitoring of Hb levels. Noninvasive techniques, including imaging and photoplethysmography (PPG) signals combined with machine learning techniques, are being investigated for continuous measurements of Hb. However, the availability of real data to train the algorithms is limited to establishing a generalization and implementation of such techniques in healthcare settings. In this work, we present a computational model based on Monte Carlo simulations that can generate multispectral PPG signals that cover a broad range of Hb levels. These signals are then used to train a Deep Learning (DL) model to estimate hemoglobin levels. Through this approach, valuable insights about the relationships between PPG signals, oxygen saturation, and Hb levels are learned by the DL model. The signals were generated by propagating a source in a volume that contains the skin tissue properties and the target physiological parameters. The source consisted of plane waves using the 660 nm and 890 nm wavelengths. A range of 6 g/dL to 18 dL Hb values was used to generate 468 PPGs to train a Convolutional Neural Network (CNN). The initial results show high accuracy in detecting low levels of Hb. To the best of our knowledge, the complexity of biological interactions involved in measuring hemoglobin levels has yet to be fully modeled. The presented model offers an alternative approach to studying the effects of changes in Hb levels on the PPGs signal morphology and its interaction with other physiological parameters that are present in the optical path of the measured signals.

Photonic decision making for solving the multi-armed bandit problem has been reported recently. However, experimental studies have been limited owing to technological difficulties. In this study, we experimentally demonstrate photonic decision making for solving a large-scale multi-armed bandit problem using optical spatiotemporal chaos generated by a semiconductor laser, spatial light modulator, and camera. We implement a parallel configuration of decision making with up to 512 slot machines, which is much larger than the previous experiment by two orders of magnitude. We found better scaling characteristics between decision-making performance and the number of slot machines than those obtained from software-based algorithms.

Brain tissue segmentation from MR images is a critical step for quantifying the brain morphology in neuroimaging studies. While deep learning (DL) based brain tissue segmentation methods have achieved promising performance, most of them are built upon supervised learning and therefore their performance is bounded by the training data used and limited by the small size of datasets with manual segmentation labels. To leverage the large amount of unlabeled brain imaging data, we develop an unsupervised DL model for joint brain tissue segmentation and bias field estimation using cascaded convolutional networks. The proposed DL model consists of multiple cascaded bias field estimation modules and one segmentation module. The bias field estimation modules are applied to the input image for estimating the bias field and generating a bias-free image recursively, and the bias field corrected image is then fed into the segmentation module to obtain the brain tissue segmentation result. A Gaussian mixture model is adopted to characterize the bias-free image with tissue-specific intensity statistics and the model fitting error is adopted as the loss function to guide the optimization of the model parameters progressively in an unsupervised setting. We have evaluated the proposed method on the HCP-Aging and HCP-Development datasets. Quantitative results have demonstrated that our unsupervised DL model could obtain competitive bias field correction and segmentation performance, compared with state-of-the-art bias field correction methods and unsupervised segmentation methods.

Evaluating disorders of consciousness (DoC) in patients typically involves traditional medical exams like the Coma Recovery Scale Revised and clinical EEG. DoC can range from coma to unresponsive wakefulness syndrome (UWS) and minimally conscious states (MCS- & MCS+). This raises the question of whether resting state fMRI and machine learning can help determine the level of consciousness in these patients. However, there are challenges with machine learning due to the scarcity of data from UWS patients and the severely unbalanced distribution of consciousness classes. Additionally, medical diagnoses may be inaccurate, particularly for UWS patients who can have covert consciousness. Our study proposes a combination of large-scale Granger causality as a biomarker and neural networks as a downstream machine learning pipeline to classify and regression problems in DoC patients. The proposed method provides a baseline for the task at hand, using causally explainable biomarkers of brain region interactions, and achieves more than 80.0% accuracy on 87 neuroimaging data with four classes of the level of unconsciousness.

Landmark detection is critical in medical imaging for accurate diagnosis and treatment of diseases. While there are many automated methods for landmark detection, the potential of transformers in this area has not been fully explored. This work proposes a transformer-based network for accurate anatomical landmark detection in chest x-ray images. By leveraging the combined power of transformers and U-Net, our method effectively captures global context and spatial dependencies, leading to robust landmark localization. The proposed method outperforms state-of-the-art methods on Chest x-ray datasets, reducing mean radial error from 5.57 to 4.68 pixels. The experiments show that the transformer-based method can effectively learn complex spatial patterns in medical images. The results of this method show the potential to improve the precision and efficiency of tasks such as surgical planning and detecting abnormalities in medical images.

Artificial intelligence and machine learning (AI/ML) are paving the way to help unravel the complexities of the human brain and shed light on the underpinnings of atypical behaviors. This is critical for early detection of mental illnesses and personalized treatment prescription. Here, key AI/ML strategies to leverage the richness of brain data are discussed, including multimodal fusion, multidimensional and deep latent representations, as well as multi-level linkage detection. Each of these is presented through a general overarching concept of subspaces. Examples from original research and the brain imaging field at large illustrate the application of each strategy to brain data from multiple modalities. Markedly, these approaches tackle challenges that extend readily to other fields outside of brain imaging.

Chromatic aberrations can significantly diminish image quality when employing multiple wavelengths for imaging with a single optical system due to dispersion in both optical system and samples. To tackle this problem, we propose a novel approach using an adaptive achromatic lens, which is controlled by a trained Reinforcement Learning agent as part of a machine learning method. Notably, our method corrects chromatic aberrations prior to the imaging process, distinguishing it from conventional software-based post-processing approaches.

Despite remarkable advances in neuromorphic photonics, nanophotonic devices enabling neuron-like spiking emission and detection—needed for spike-based computation—are still at an early stage of development. Here, we present recent achievements on a photonic neuro-architecture based on III–V integrated spiking circuit nodes using nanoscale quantum resonant tunneling diodes (nanoRTDs) with negative differential resistance. We discuss the integration of nanoRTDs with nano-photodetectors (nanoPDs), nano-light-emitting diodes (nanoLEDs) and other light sources (e.g., VCSELs), to realize on-chip sensory neuron emitter and receiver spiking nodes. The RTD-based nanophotonic neuron devices and circuits for sensing and spike processing multi-functionalities pave the way for energy-efficient signal processing and sensory neuromorphic systems in edge computing AI applications.

Machine learning (ML) is becoming a ubiquitous and powerful tool helping to address challenges in countless fields. Applications of ML addressing optics challenges have been extensively studied in recent years opening up new research directions. In particular, here, we review some of our current efforts and provide examples of successful applications of ML to the characterization of photonic devices, design, and modeling of optical subsystems, and complete end-to-end optical system optimization. ML and statistical tools can yield additional insight from measurement data, e.g. by targeted filtering of noise sources. They have also been shown to assist complex or inaccurate physics-based models through black and grey-box modeling of photonics components or subsystems. Such ML-aided models have enabled easier optimization and design (including inverse design) of optical systems.

We demonstrate a multiplexed fluorescent vertical flow assay (fxVFA) processed by a hand-held reader and a deep learning-based algorithm for quantification of three biomarkers, i.e. myoglobin, creatine kinase-MB (CK-MB) and heart-type fatty acid binding protein (FABP) from human serum samples. fxVFA operation takes <15 min and requires 50 µL of serum sample. fxVFA achieved <0.52 ng/mL limits-of-detection for all three analytes with minimal cross-reactivity between the antigens. Furthermore, quantification performance of fxVFA was tested on 16 serum samples and fxVFA-predicted concentrations had >0.9 coefficients of determination and <15 % coefficients of variation with the respect to a standard ELISA test.

The explainable artificial intelligence (AI) techniques can help us explain the predictive behavior of a machine learning model without using the information of the model architecture, model parameters, or the training strategies but using the human (users) interpretable features and the predictions. This paper presents a technique, namely nEGXAI, to develop explainable AI models. The current explainable AI approaches are generally built on the affirmation-based logic. In other words, when a model predicts a class label, the explainable AI may explain the prediction results by the features that are associated with the predicted class and interpretable by the users. In contrast to the use of affirmation-based logic, the proposed approach uses the negation-based mathematical logic. In other words, when a model predicts a class label, the proposed nEGXAI can explain why the predicted class is not one of the other classes by using the features that are not associated with the predicted class. The proposed nEGXAI is also a model-agnostic technique in the sense that it does not assume any knowledge of model-related information but uses only the features from the data and the predictions, the same way the Local Interpretable Model-agnostic Explanations (LIME) and SHapley Additive exPlanations (SHAP) techniques use. The nEGXAI utilizes the mathematical concept of constructive negation and its effect on the spectral density function of the features in the Fourier domain. It allows nEGXAI to capture the possible prediction failures that are human interpretable and explainable; hence, it delivers a trustworthy explainable AI for decision making.

We have applied physics aware training (PAT) to diffractive deep neural networks (D2NN) consisting of multiple spatial light modulators (SLMs) to close the reality gap between the simulation model and the physical system. Compared to conventional training methods using only simulation models, PAT improves classification accuracy in the experiment. In this method, an analytic expression for backpropagation is based on Rayleigh-Sommerfeld diffraction integral as conventional, but the backpropagated error values are replaced by the measured values.

With the technological advancements in machine learning, it has become more prevalent to use learning techniques for clinical decision-making based on medical images. One of the state-of-the-art methods used for this purpose is Convolutional Neural Networks (CNN) for medical image segmentation and deep learning models for disease detection and classification. In this paper, we propose a framework for image segmentation using hierarchical CNNs to classify different types of cells using small frame images. This paper aims to generalize the segmentation of cancer cells, starting with cervix cancer. The first step of the framework is to achieve automatic nucleus and cell masking of the images using U-Net. The images are then segmented into “satisfactory” and “unsatisfactory” categories to determine whether these images can be used in our classification model. Using the hierarchical CNN, the satisfactory images are clustered based on cell types since the cell features that need to be considered vary between different cell types. Lastly, our classification model is trained with automatically segmented images to classify different cancer types based on cell images using various features, such as the area of the nucleus, the ratio of the nucleus area and cytoplasm area and the visual morphology of chromatin strands in the nucleus. To demonstrate the performance of the proposed framework, a labeled dataset, taken from the Detay Pathology and Cytology Laboratory, with over 100 images were used.