This PDF file contains the front matter associated with SPIE Proceedings Volume 12227, including the Title Page, Copyright information, Table of Contents, and Conference Committee listings.

We present a study to address the reconstruction of synthetic aperture radar (SAR) images using machine learning. From previous work, we utilize a single, fully-connected layer to learn the sensing matrix of the forward scattering problem by using a set of reflectivites as the inputs and the SAR measurements as the outputs. Using this result, we estimate the reflectivity of the SAR measurements by applying the conjugate transpose of the learned sensing matrix to the SAR measurement data. Then we further improve the reconstructions of the reflectivity using convolutional layers. Employing a training set made up of 50,000 images of randomly placed point scatterers as the reflectivities, we simulate SAR measurement data using a physical model for the sensing matrix. We apply the learned sensing matrix to our SAR measurement data and use this estimate of the reflectivity as the inputs to the model, while the true reflectivities are the outputs. The model is trained to reconstruct images containing a single target. We find that the resulting reconstructions are sharper images than those from the initial estimate from applying the conjugate transpose of the learned sensing matrix. In particular, we find that the background noise is significantly decreased. In addition, we test this model on a different dataset with multiple targets as reflectivities. Similar to previous results, and with no additional training, the model applied to data with multiple targets also demonstrated improved reconstructions of reflectivities.

Current changes in climate conditions are causing rapid reductions in Arctic sea ice. This loss of sea ice has environmental implications as well as economic and global security implications, especially in regards to Arctic navigation. Rapid changes in sea ice influence when an Arctic route can become navigable, and thus understanding such changes is crucial for operation planning and routing. The Arctic’s spatial extent and severe environment are well suited for observation via satellite remote sensing, and the problem space lends itself well to a supervised machine learning approach. However, such an approach is limited by the lack of labeled sea ice data sets with both the spatial and temporal density and resolution to pair with image data. In this study, we develop methods to derive sea ice labels directly from satellite synthetic aperture radar (SAR) data. We use single-look complex data from the Sentinel-1 constellation collected in the Extra-Wide swath mode, which has optimal imaging parameters for sea ice observation. We expand on existing methods of deriving labels from SAR data using H-α plane polarimetric classification techniques by examining additional polarimetric parameters. We then develop new classification rules by training a decision tree classifier using a labeled data set made available by the National Snow & Ice Data Center. We focus our analysis on data collected in Summer of 2020 and Summer and Fall of 2021 covering the Greenland and Barents Seas, Central Arctic Ocean, and Baffin Bay.

Additive manufacturing (AM) is rapidly expanding as a result of its ability to efficiently produce complex parts that are difficult or impossible to fabricate with traditional manufacturing processes. However, a key factor currently limiting wide- scale implementation of this technology is the lack of confidence in the reliability of AM components. Laser powder bed fusion (LPBF) is an AM process where lasers are used to successively melt layers of metal powder to produce threedimensional parts. Selecting the appropriate processing parameters, such as laser power and scan speed, directly impacts the quality and performance of AM depositions. By studying the characteristics of single-track melt pools produced via LPBF, the appropriate processing parameters for a given build may be determined. Quantification of critical melt pool characteristics – such as width, depth, and surface morphology – is primarily performed through manual measurements on high resolution images; a process which can be immensely time consuming and is subject to user bias. In this study, computer vision techniques, such as edge detection and texture segmentation, are used to automatically detect and quantify these characteristics. Clustering methods were then applied to these measurements to classify the processing regime for the melt pool geometry and associated operating parameters used in its creation. Operators can then use the classification information to select build parameters which reduce the likelihood of defects in constructed parts. This characterization process has been demonstrated through automated classification of single-tracks experiments for 4340 steel produced by an EOS M290 LPBF system.

This paper presents SAFE, a prototype system for supporting the fish landings control of small-scale fishing boats in Chile. SAFE is a modern solution for fishery inspection that automatically discriminates fish species using machine learning. Here, we present a version of SAFE that classifies five target pelagic fish species in Chile: anchovy, Chilean jack mackerel, hake, mote sculpin, and sardine. The system has two stages; the first detects and segments all fish appearing in an image. These segmented images then feed the second stage, which perform species classification. A database of approximately 266 images from these five fish species was constructed for training, validation, and testing purposes. For the fish detection stage, we exploited transfer learning to train Mask R-CNN architectures, an instance segmentation model. As for the fish species classification stage, we exploited transfer learning to train ResNet50 and VGG16 deep learning architectures. Results show that SAFE achieves between 90% and 96.3% macro-average precision (MP) when classifying the five fish species mentioned above. The best architecture, composed of a Mask R-CNN-based detector and a VGG16-based classifier, achieves an MP of 96.3%, which could process a single fish as quick as 16.67 FPS, and one whole 1920x1080-pixel image as quick as 2 FPS.

The maximum classifier discrepancy method has achieved great success in solving unsupervised domain adaptation tasks for image classification in recent years. Its basic structure consists of a feature generator and two classifiers that aim to maximize the classifier discrepancy while minimizing the generator discrepancy of the target samples. This method improves the performance of the existing adversarial training methods by employing task-specific classifiers that remove the ambiguity in classifying the target samples near the class boundaries. In this paper, we propose a modified network architecture and two training objectives to further boost the performance of the maximum classifier discrepancy method. The first training objective minimizes the feature level discrepancy and forces the generator to generate domain invariant features. This training objective is particularly beneficial when the source and the target domain distributions are vastly different. The second training objective that works at the mini-batch level aims at creating a uniform distribution of the target class predictions by maximizing the entropy of the expectation of the target class predictions. We show through extensive empirical evaluations that the proposed architecture and training objectives significantly improve the performance of the original algorithm. Furthermore, this method also outperforms the state-of-the-art techniques in most unsupervised domain adaptation tasks.

In this paper, we investigate the feasibility of using machine learning methods for the estimation of StructuralThermal-Optical-Performance (STOP) models of reflective optics. We use a model of a Newtonian telescope system to test machine learning methods. To generate the estimation data, we model and simulate a transient finite-element STOP model of the Newtonian telescope by using COMSOL Multiphysics and LiveLink for MATLAB software module. We use a feedforward neural network structure to estimate the STOP model. The inputs and outputs of the neural network correspond to the inputs and outputs of a Vector AutoRegressive eXogenous (VARX) model. Our results show that large-scale STOP dynamics can be effectively approximated by a loworder neural network model. Consequently, low-order VARX or state-space models can be reconstructed from the parameters of the estimated feedforward neural network, and used for the prediction, state estimation, and design of model-based controllers. We use the TensorFlow and Keras machine learning libraries and Python to estimate the feedforward neural network model. The developed COMSOL, MATLAB, and Python codes are available online.

Nowadays, sophisticated ray-tracing software packages are used for the design of optical systems, including local and global optimization algorithms. Nevertheless, the design process is still time-consuming with many manual steps, and it can take days or even weeks until an optical design is finished. To address this shortcoming, artificial intelligence, especially reinforcement learning, is employed to support the optical designer. In this work, different use cases are presented, in which reinforcement learning agents are trained to optimize a lens system. Besides the possibility of bending lenses to reduce spherical aberration, the movement of lenses to optimize the lens positions for a varifocal lens system is shown. Finally, the optimization of lens surface curvatures and distances between lenses are analyzed. For a predefined Cooke Triplet, an agent can choose the curvature of the different surfaces as optimization parameters. The chosen surfaces and the distances between the lenses will then be optimized with a least-squares optimizer1 . It is shown, that for a Cooke Triplet, setting all surfaces as variables is a good suggestion for most systems if the runtime is not an issue. Taking the runtime into account, the selected number of variable surfaces decreases. For optical systems with a large number of degrees of freedom an intelligent selection of optimization variables can probably be a powerful tool for an efficient and time-saving optimization.

A popular framework to design optical components is inverse design. Here, one starts from a desired optical response and tries to find an optical design with exactly this response. In recent years, Machine Learning and in particular Deep Learning have been used in this regard. Many successes have been reported in the literature. What often arises in practice is the non-uniqueness problem. Multiple designs can lead to similar or even exactly the same optical responses. This is a complication, because most methods can only propose one design. The non-uniqueness problem can be mitigated by considering the degrees of freedom of the system. While multiple geometric parameters can correspond to the same response, the degrees of freedom define them uniquely. When working with these parameters, there is a one-to-one mapping between design and response. In order to find the degrees of freedom in an optical system, we propose a neural network architecture called a beta-VAE. We study its behaviour for analytical datasets of sine waves and damped harmonic oscillators. This gives us insight into the working of the neural network. We find that degrees of freedom can be retrieved for physical systems relevant to Optics.

As more and more cameras are used for machine perception, the optical design process still relies on key indicators such as point spread function (PSF), modulated transfer unction (MTF) based on aberration minimization. This process has proven efficient for human vision but is not tailored for machine perception. Given a specific computer vision task, it is not always necessary to target the same key performance indicators (KPIs) than when images are visualized by humans. Moreover, this image quality might change during a camera lifespan with the appearance of defocus for example. It is crucial to be able to determine how this kind of degradation can affect a computer vision task. In this work we study the impact of defocus on 2D object identification and show that, for a certain design, it is not impacted by image degradation under a certain threshold. We also demonstrate that this threshold is higher for lower f-number which makes them better design candidates.

Currently, the external quantum efficiency (EQE) of OLED devices is almost at a limit. One of the causes of these limitations is that the performance of phosphorescent molecules designed by humans is at the limit. To overcome the limitation, we introduce a methodology for screening multiple molecules using reinforcement learning and density function theory (DFT). We first set up an environment where phosphorescent molecules can be designed, and explore molecular space using reinforcement learning and density functional tight binding (DFTB) theory. Subsequently, the molecules produced as candidate groups are screened by wavelength, bond dissociation energy (BDE), photoluminescence quantum yield (PLQY), and emitting dipole orientation (EDO) using the time-dependent DFT and molecular dynamics (MD). As a result, we successively found new ways to design new green cores.

A lightweight deep learning neural network model for predicting the refractive index structure constant (C²_n) was constructed using atmospheric data collected at the TISTEF laser range over a 90 day period. The model used variables such as barometric pressure, wind speed, wind direction, air temperature, corrected average ground temperature, relative humidity, solar kilowatts, solar mega joules, and various derived variables to make its predictions. The model was evaluated using measurements recorded from a Scintec BLS 900. The predictions from the model were highly correlated with the measured values of C²_n.

Robustness to image quality degradations is critical for developing deep neural networks for real-world image classification. Previous efforts have pursued robustness by exploring how various types of blur, noise, contrast, compression, color, etc. degrade image quality and impact image classification performance. This paper extends this discussion to include optical aberrations, which are fundamental to the lens design of imaging systems and enable further discussion of DNN performance in the context of hardware design. In this paper, multiple state-of-the-art DNN models are evaluated for their image classification performance with imagery that has been degraded by various optical aberrations.

Manual optimization of experimental parameters can quickly become too complex and time-consuming if more than a few correlated parameters need to be adjusted. We discuss automating this process using Bayesian optimization. This machine learning-based method is particularly suitable because it can handle noisy measurements, performs a global search and requires relatively few experimental runs. We discuss the efficient, scalable implementation of Bayesian optimization, present practical applications for tuning experimental parameters, and compare it with other local and global heuristic methods to show its application range.

In the production of display screen modules, multi-faceted quality control is performed. One of the processes is detection of defects on and between module components such as particles, scratches and air bubbles using a 3D optical microscope. Technicians view a stack of images of potential defect areas and make a qualitative assessment of the sample. However, this is made difficult by the artifacts in the unfocused image layers. Moreover, there is a large discrepancy in the detection tendencies of the technicians. In order to standardize and automate the classification of major and minor defects in products, we propose a convolutional neural network based binary classification that makes use of the normal angle and oblique angle images. The decision factors affecting the classification of the sample include defect position, size, and shape. In order to reflect these factors, the microscopic images of the sample are taken in varying focal depths from normal and oblique angles. Then, the maximum intensity projection (MaxIP) and minimum intensity projection (MinIP) in the xy, yz, xz plane are created. The set of MaxIP and MinIP are used to train a modified VGG-network. Each plane differs in size, so MaxIP and MinIP of every plane was independently added as input to the network and were concatenated in the fully connected layer. Being that the dataset used for this work composed of 185 major defect samples and 2036 minor defect samples, augmentation was essential. In order to even out the major and minor defect sample ratio, random affine transformation was performed on the major defect sample images. The proposed method of binary classification performs with a total accuracy of 98.6%.

Fitting a theoretical model to experimental data in a Bayesian manner using Markov chain Monte Carlo typically requires one to evaluate the model thousands (or millions) of times. When the model is a slow-to-compute physics simulation, Bayesian model fitting becomes infeasible. To remedy this, a second statistical model that predicts the simulation output – an “emulator” – can be used in lieu of the full simulation during model fitting. A typical emulator of choice is the Gaussian process (GP), a flexible, non-linear model that provides both a predictive mean and variance at each input point. Gaussian process regression works well for small amounts of training data (n < 10³), but becomes slow to train and use for prediction when the data set size becomes large. Various methods can be used to speed up the Gaussian process in the medium-to-large data set regime (n < 10⁵), trading away predictive accuracy for drastically reduced runtime. This work examines the accuracy-runtime trade-off of several approximate Gaussian process models – the sparse variational GP, stochastic variational GP, and deep kernel learned GP – when emulating the predictions of density functional theory (DFT) models. Additionally, we use the emulators to calibrate, in a Bayesian manner, the DFT model parameters using observed data, resolving the computational barrier imposed by the data set size, and compare calibration results to previous work. The utility of these calibrated DFT models is to make predictions, based on observed data, about the properties of experimentally unobserved nuclides of interest e.g. super-heavy nuclei.

High-fidelity physics simulations, such as the Energy Exascale Earth System Model (E3SM), are generating ever increasing quantities of data. In the near future, it will be infeasible to store the full computation after the simulation. To facilitate the training of complex machine learning models in the future, there is a glaring need for in-situ inference algorithms. In this work, we focus on the need for spatio-temporal inference models which are (i) highly scalable, (ii) distributed, (iii) able to capture small-scale, high-resolution structures, and (iv) longrange global structure. One possible approach is to leverage algorithms for fast approximate Gaussian process regression, such as the sparse variational Gaussian process (SVGP). In this work, we present a Hierarchical SVGP (HSVGP) model in which local models capture small-scale structures in-situ, while large-scale structures are captured and communicated by a global SVGP. We also consider the simpler approach in which independent, parallel sparse Gaussian processes are fit to each local region with no additional global information. Using both synthetic data and E3SM model output, we show the success of the HSVGP approach in some settings, but also the surprisingly good performance of the independent approach.

We describe the methods for automating the workflow for rapidly measuring and producing elemental maps of large-area samples using the Submicron Resolution X-ray Spectroscopy Beamline (SRX) at the National Synchrotron Light Source II, Brookhaven National Laboratory, through a novel combination of supervised (support vector machine) and unsupervised (cluster analysis) machine learning algorithms. SRX has the capability to create centimeter area full spectrum x-ray fluorescence (XRF) maps non-destructively with special detector and beam configurations. To facilitate the automation of this process, we discuss the development of the Synchrotron Network Automation Program in Python (SnapPy) software package that automates measurements such that SnapPy will control everything from beamline machine control to data acquisition and analysis. The only intervention that will need to be performed by beamline staff will be to physically install and remove samples. This will allow us to run measurements overnight or during times when beamline staff would not otherwise be available.

Due to recent advancements in low-temperature detectors for x-ray emission spectroscopy (XES), we are now able to generate higher resolution spectra with larger bandwidth than previously attainable. While this more accurate detection technology will allow for more reliable sample classification, current analysis is insufficient to handle the new level of detail. The limitations of current classification methods is exacerbated by the difficulty producing unifying theory for the physical mechanisms responsible for x-ray emission. Currently, it is well understood how to detect elemental concentrations in a sample, but not the chemical composition. To address this gap in the data analysis, we use a domain-knowledge naive approach to data mining to discover regions-of-interest (ROIs) relevant to oxidation state of chemical compounds. Typical features of the spectra in the found ROIs, such as relative peak heights and location, are used in highly transparent machine learning methods. We also train a convolutional neural network (CNN) using Monte-Carlo derived data based on a recorded training set of spectra. Both methods are applied to a testing set of the same compounds measures a year prior. Our results indicate that, while the ROIs found using the naive mining approach potentially carry information regarding oxidation state, the signatures are more subtle than relative peak heights and positions. This is supported by saliency maps of the CNN when classifying the test set. These methods are the first step to gaining the understanding of XES required to use this detection technology in a real-time chemical imaging application.

Machine learning (ML) based compact device modeling provides the opportunity for process-aware device modeling and thus process-aware circuit simulation. In contrast, incorporating semiconductor manufacturing parameters into compact models (CM) and subsequent circuit simulations using pure physics-based CM is difficult. We demonstrate process-aware circuit simulation where the effect of plasma treatment and thermal annealing can be directly reflected on the circuit output in SPICE transient simulation. The Verilog-A model input is V, frequency(f), area(A), and process conditions, i.e., plasma surface treatment (PST) and post-metal annealing (PMA). The MOSCAP capacitance-voltage (CV) characteristics under illumination are described by ML compact models.

We previously demonstrated a machine learning based regions-of-interest (ROI) finding tool for X-ray fluorescence microscopy, called XRF-ROI-Finder at the 9-ID beamline in Argonne National Laboratory.1 Bacterial cell treatment type prediction and recommendation for steering experiments were performed via the application of fuzzy k-means clustering algorithm. ROI-Finder takes the fluorescence microscopy images, performs segmentation and detects individual E.coli cells, extracts features for principal component analysis, and ultimately performs label-free clustering for cell treatment type prediction and recommendation for similar cells to perform automatic steering experimentation. In this paper, we assess two additional clustering method, namely hierarchical agglomerative clustering (HAC) and density based spatial clustering of applications with noise (DBSCAN) algorithm. The ROI-Finder software is hosted at https://github.com/aisteer/ROI-Finder.

Alzheimer’s is a chronic disease that impairs an individual’s cognitive ability to carry out daily actions. According to the World Health Organization, Alzheimer’s and other dementias rank 7th in the top causes of death worldwide. The US Centers for Disease Control (CDC) estimates that the number of Americans with Alzheimer’s will triple and reach 14 million in 2060. Our research explores Multi-Modal AI ensemble methods that combine MRI (Magnetic Resonance Imaging) scans with image-featurized brain volumetrics (Freesurfer) and clinician notes to effectively forecast Alzheimer’s. We used the OASIS3 dataset, a longitudinal study which tracked 1098 patients over thirty years and generated a total of 2168 MR sessions and over 6000 clinician visit notes. We evaluated the Convolutional Neural Networks (CNN) ResNet101, ResNet152 and ResNet200 for forecasting Alzheimer’s from the MRI scans. We then combined the MRI forecasts with Machine Learning techniques (Random Forest and K Nearest Neighbors) for forecasts based on Freesurfer featurized brain volumetrics and additional clinical data. The individual forecasts were combined using four ensemble methods (majority voting, aggressive which required only one modality, conservative which required all three and conditional where clinician notes took priority) to forecast an individual’s eventual Clinical Dementia Rating (CDR). The final results showed that a patient’s Alzheimer’s CDR can be forecasted early with an accuracy of over 94%, and that harmful False Negatives can be reduced by 2x-15x depending on the ensemble method chosen. These methods can assist in the early treatment of susceptible patients.

Automatic classification of child facial expressions is challenging due to the scarcity of image samples with annotations. Transfer learning of deep convolutional neural networks (CNNs), pretrained on adult facial expressions, can be effectively finetuned for child facial expression classification using limited facial images of children. Recent work inspired by facial age estimation and age-invariant face recognition proposes a fusion of facial landmark features with deep representation learning to augment facial expression classification performance. We hypothesize that deep transfer learning of child facial expressions may also benefit from fusing facial landmark features. Our proposed model architecture integrates two input branches: a CNN branch for image feature extraction and a fully connected branch for processing landmark-based features. The model-derived features of these two branches are concatenated into a latent feature vector for downstream expression classification. The architecture is trained on an adult facial expression classification task. Then, the trained model is finetuned to perform child facial expression classification. The combined feature fusion and transfer learning approach is compared against multiple models: training on adult expressions only (adult baseline), child expression only (child baseline), and transfer learning from adult to child data. We also evaluate the classification performance of feature fusion without transfer learning on model performance. Training on child data, we find that feature fusion improves the 10-fold cross validation mean accuracy from 80.32% to 83.72% with similar variance. Proposed fine-tuning with landmark feature fusion of child expressions yields the best mean accuracy of 85.14%, a more than 30% improvement over the adult baseline and nearly 5% improvement over the child baseline.

According to the Alzheimer’s Association, roughly one in nine people age 65 and older suffer from the burden of Alzheimer’s dementia and one in three seniors dies with Alzheimer’s or another dementia. Recent advances have been made in early diagnosis of Alzheimer’s Disease including utilizing machine learning techniques to identify abnormalities associated with Alzheimer’s Disease in magnetic resonance imaging (MRI) data. In this paper, we will explore how the pre-processing of two-dimensional (2D) slices of MRI data using digital signal processing techniques affect a machine learning classifier. This work differs from other studies as it focuses on the methods used to pre-process the MRI data to highlight abnormalities rather than optimizing the machine learning approach based on the available data provided. We show that employing digital signal processing techniques, specifically low-pass and high-pass filtering, to 2D slices of an MRI in the frequency domain can improve the performance of a basic machine learning classifier. This is a promising result which has the potential to improve the performance of state-of-the-art machine learning classifiers simply by pre-processing the data using a digital signal/image filter.

Vision-based detection and classification systems have a great potential in supporting agricultural processes, such as monitoring fields or weed management. They build the foundation for automated weeding solutions that have the potential to reduce the amount of chemicals that have wide-ranging environmental effects. Unmanned aerial vehicles (UAVs) have the potential to operate on wet soils, whereas conventional agricultural robots only have a limited mobility in this specific environment. To operate in a UAV-based system, the vision-based detection system has to meet specific requirements in regards to the power consumption, overall detection performance, real-time capability, weight and size. This paper evaluates early research results of a vision-based deep learning approach for weed detection in horticulture, more specifically arboriculture. To our knowledge, this research has not been conducted before. The vision-based deep learning system is evaluated with images from horticultures under normal commercial growing conditions using different edge computing devices that fit the weight, size, and power-consumption requirements of a UAV-based application. These results are benchmarked against a high-performance PC-GPU-configuration. The experimental results are presented in the paper.

For several years, Computer-Aided Detection (CAD) systems have been used by radiologists as a second interpreter for breast cancer detection in digital mammography. However, for every true-positive cancer detected by a CAD system, more false predictions must be revised by the expert to avoid an unnecessary biopsy. On the other hand, the research community has been exploring different approaches for the detection and classification of breast abnormalities. Machine learning, and particularly deep learning approaches, are being used to analyze digital mammography images. Nevertheless, most of the models proposed so far are trained on a single database and do not have high reliability. In this work, several deep learning models were compared for benign-malign mammography classification. A pre-processing stage is designed to remove noise and extract features using local image phase information. Then, a machine learning approach is utilized for digital mammography classification. Experimental results are presented using various digital mammography datasets and evaluated under different performance metrics.

Image classification forms an important class of problems in machine learning and is widely used in many realworld applications, such as medicine, ecology, astronomy, and defense. Convolutional neural networks (CNNs) are machine learning techniques designed for inputs with grid structures, e.g., images, whose features are spatially correlated. As such, CNNs have been demonstrated to be highly effective approaches for many image classification problems and have consistently outperformed other approaches in many image classification and object detection competitions. A particular challenge involved in using machine learning for classifying images is measurement data loss in the form of missing pixels, which occurs in settings where scene occlusions are present or where the photodetectors in the imaging system are partially damaged. In such cases, the performance of CNN models tends to deteriorate or becomes unreliable even when the perturbations to the input image are small. In this work, we investigate techniques for improving the performance of CNN models for image classification with missing data. In particular, we explore training on a variety of data alterations that mimic data loss for producing more robust classifiers. By optimizing the categorical cross-entropy loss function, we demonstrate through numerical experiments on the MNIST dataset that training with these synthetic alterations can enhance the classification accuracy of our CNN models.

While they have similar symptoms, Psoriasis and Eczema have vastly different underlying causes and behaviors. Our research explores state of the art Deep Learning techniques (ResNet152 and MobileNetV2) for distinguishing Psoriasis and Eczema. The dataset includes 99 images of hand Psoriasis, 102 hand Eczema, 49 foot Psoriasis, and 42 foot Eczema. Using ResNet152 and MobileNetV2 in a TensorFlow prototype, and hyper-parameter tuning both across a range of parameters, we demonstrated that both architectures deliver substantial accuracy improvements over past research. MobileNetV2 foot samples had a predictive accuracy of 94.11%, ResNet152 foot samples had a highest accuracy of 100%. The two deep learning solutions also distinguished conditions for hand images with accuracies ranging from 73.68-89.47%. The results are optimistic for using deep learning with skin images for diagnostic assistance for these two conditions. For future work, we are exploring image featurization techniques in combination with both neural networks and statistical machine learning techniques to further improve predictive performance.

Quantum machine learning by superposition and entanglement, for disease categorization utilizing OCT images will be discussed in this paper. To the best of our knowledge, this is the first application of the quantum computing element in a neural network model for classifying ophthalmological disease. The model was built and tested with PennyLane (PennyLane.ai), an open-source software tool based on the concept of quantum differentiable programming. The model training circuit functioning was tested on an IBM 5 qubits system “ibmq_belem” and a 32 qubits simulator “ibmq. qasm_simulator”. A hybrid quantum and classical model with a 2 qubit QNode converted layer with operations such as Angle Embedding, BasicEntanglerLayers and measurements were the internal operations of the qlayer. Drusen, Choroidal neovascularization (CNV), and Diabetic macular edema (DME) OCT images formed the abnormal/disease class. The model was trained using 414 normal and 504 abnormal labelled OCT scans and the validation used 97 and 205 OCT scans. The resulting model had an accuracy of 0.95 in this preliminary 2-class classifier. This study aims to develop a 4-class classifier with 4 qubits and explore the potential of quantum computing for disease categorization. A preliminary performance analysis of quantum Machine Learning, the steps involved, and operational details will be discussed.