High Classification Accuracy Research Articles

A Transformer Approach for Cognitive Impairment Classification and Prediction.

Early classification and prediction of Alzheimer disease (AD) and amnestic mild cognitive impairment (aMCI) with noninvasive approaches is a long-standing challenge. This challenge is further exacerbated by the sparsity of data needed for modeling. Deep learning methods offer a novel method to help address these challenging multiclass classification and prediction problems. We analyzed 3 target feature-sets from the National Alzheimer's Coordinating Center (NACC) dataset: (1) neuropsychological (cognitive) data; (2) patient health history data; and (3) the combination of both sets. We used a masked Transformer-encoder without further feature selection to classify the samples on cognitive status (no cognitive impairment, aMCI, AD)-dynamically ignoring unavailable features. We then fine-tuned the model to predict the participants' future diagnosis in 1 to 3 years. We analyzed the sensitivity of the model to input features via Feature Permutation Importance. We demonstrated (1) the masked Transformer-encoder was able to perform prediction with sparse input data; (2) high multiclass current cognitive status classification accuracy (87% control, 79% aMCI, 89% AD); (3) acceptable results for 1- to 3-year multiclass future cognitive status prediction (83% control, 77% aMCI, 91% AD). The flexibility of our methods in handling inconsistent data provides a new venue for the analysis of cognitive status data.

Self-supervised motor imagery EEG recognition model based on 1-D MTCNN-LSTM network

Objective. Aiming for the research on the brain–computer interface (BCI), it is crucial to design a MI-EEG recognition model, possessing a high classification accuracy and strong generalization ability, and not relying on a large number of labeled training samples. Approach. In this paper, we propose a self-supervised MI-EEG recognition method based on self-supervised learning with one-dimensional multi-task convolutional neural networks and long short-term memory (1-D MTCNN-LSTM). The model is divided into two stages: signal transform identification stage and pattern recognition stage. In the signal transform recognition phase, the signal transform dataset is recognized by the upstream 1-D MTCNN-LSTM network model. Subsequently, the backbone network from the signal transform identification phase is transferred to the pattern recognition phase. Then, it is fine-tuned using a trace amount of labeled data to finally obtain the motion recognition model. Main results. The upstream stage of this study achieves more than 95% recognition accuracy for EEG signal transforms, up to 100%. For MI-EEG pattern recognition, the model obtained recognition accuracies of 82.04% and 87.14% with F1 scores of 0.7856 and 0.839 on the datasets of BCIC-IV-2b and BCIC-IV-2a. Significance. The improved accuracy proves the superiority of the proposed method. It is prospected to be a method for accurate classification of MI-EEG in the BCI system.

Single-cell classification based on label-free high-resolution optical data of cell adhesion kinetics.

Selecting and isolating various cell types is a critical procedure in many applications, including immune therapy, regenerative medicine, and cancer research. Usually, these selection processes involve some labeling or another invasive step potentially affecting cellular functionality or damaging the cell. In the current proof of principle study, we first introduce an optical biosensor-based method capable of classification between healthy and numerous cancerous cell types in a label-free setup. We present high classification accuracy based on the monitored single-cell adhesion kinetic signals. We developed a high-throughput data processing pipeline to build a benchmark database of ~ 4500 single-cell adhesion measurements of a normal preosteoblast (MC3T3-E1) and various cancer (HeLa, LCLC-103H, MDA-MB-231, MCF-7) cell types. Several datasets were used with different cell-type selections to test the performance of deep learning-based classification models, reaching above70-80% depending on the classification task. Beyond testing these models, we aimed to draw interpretable biological insights from their results; thus, we applied a deep neural network visualization method (grad-CAM) to reveal the basis on which these complex models made their decisions. Our proof-of-concept work demonstrated the success of a deep neural network using merely label-free adhesion kinetic data to classify single mammalian cells into different cell types. We propose our method for label-free single-cell profiling and in vitro cancer research involving adhesion. The employed label-free measurement is noninvasive and does not affect cellular functionality. Therefore, it could also be adapted for applications where the selected cells need further processing, such as immune therapy and regenerative medicine.

An intelligent fault diagnosis method for rolling bearing using motor stator current signals

Abstract In the diagnosis of rolling bearing faults, the Motor Current Signature Analysis (MCSA) method offers advantages such as low cost, simplicity, and convenience compared with using vibration signals, temperature information, and other diagnostic objects. However, owing to the interference of high-frequency noise, power frequency, and its harmonics in current signals, which can severely affect the accuracy of bearing fault diagnosis, it is extremely challenging to directly use the original current signals during bearing faults for diagnostic purposes. Therefore, this paper proposes an intelligent fault diagnosis method based on the feature reconstruction (FR) method and Convolutional Neural Networks (CNN). This method can achieve high-precision fault diagnosis using single-phase stator current signals from motors as the diagnostic objects. First, the feature reconstruction method effectively removes the impact of high-frequency noise, supply frequency, and its harmonics from the current signals, while also highlighting subtle fault feature signals to a certain extent. Second, a CNN suitable for learning the characteristics of the current signals was constructed. Through feature extraction, learning, and classification of the current signal samples processed by the feature reconstruction method, a diagnostic method with a high classification accuracy was obtained. Visualization techniques were used to intuitively present the final diagnosis results. The experimental results demonstrated the highest and average diagnostic accuracies of the proposed method in diagnosing rolling bearing fault types, with an average diagnostic accuracy of approximately 99% for actual faulty bearing samples.

Accurate wire rope defect MFL detection using improved Hilbert transform and LSTM neural network

ABSTRACT Wire rope is playing a significant role in various engineering fields. Although a great number of non-destructive testing methods are applied to wire rope defect detection, the results are greatly influenced by the signal noises and the quantitative defect detection of wire rope is consequently limited. Therefore, a new accurate and quantitative wire rope defect magnetic flux leakage (MFL) recognition method based on improved Hilbert transform and long short-term memory (LSTM) neural network is proposed. The theoretical background of the proposed signal processing models and preliminary results, as well as the influence of the main model parameters are analyzed first. Then, different wire rope defect classification principles through LSTM neural network and performance evaluation based on the improved signal processing and machine learning combined method are presented. Finally, the defect inspection and classification comparison results between the proposed method and other deep learning neural networks and pattern recognition strategies are given, which manifest that the proposed method has higher classification accuracy and shorter running time for five different wire rope defects under various servicing conditions, and is promising in accurate defect inspection for multi-scenario wire rope. Additionally, the main conclusions, disadvantages of the proposed method are summarized.

A Biologically Inspired Movement Recognition System with Spiking Neural Networks for Ambient Assisted Living Applications

This study presents a novel solution for ambient assisted living (AAL) applications that utilizes spiking neural networks (SNNs) and reconfigurable neuromorphic processors. As demographic shifts result in an increased need for eldercare, due to a large elderly population that favors independence, there is a pressing need for efficient solutions. Traditional deep neural networks (DNNs) are typically energy-intensive and computationally demanding. In contrast, this study turns to SNNs, which are more energy-efficient and mimic biological neural processes, offering a viable alternative to DNNs. We propose asynchronous cellular automaton-based neurons (ACANs), which stand out for their hardware-efficient design and ability to reproduce complex neural behaviors. By utilizing the remote supervised method (ReSuMe), this study improves spike train learning efficiency in SNNs. We apply this to movement recognition in an elderly population, using motion capture data. Our results highlight a high classification accuracy of 83.4%, demonstrating the approach’s efficacy in precise movement activity classification. This method’s significant advantage lies in its potential for real-time, energy-efficient processing in AAL environments. Our findings not only demonstrate SNNs’ superiority over conventional DNNs in computational efficiency but also pave the way for practical neuromorphic computing applications in eldercare.

Multi-purpose RNA language modelling with motif-aware pretraining and type-guided fine-tuning

AbstractPretrained language models have shown promise in analysing nucleotide sequences, yet a versatile model excelling across diverse tasks with a single pretrained weight set remains elusive. Here we introduce RNAErnie, an RNA-focused pretrained model built upon the transformer architecture, employing two simple yet effective strategies. First, RNAErnie enhances pretraining by incorporating RNA motifs as biological priors and introducing motif-level random masking in addition to masked language modelling at base/subsequence levels. It also tokenizes RNA types (for example, miRNA, lnRNA) as stop words, appending them to sequences during pretraining. Second, subject to out-of-distribution tasks with RNA sequences not seen during the pretraining phase, RNAErnie proposes a type-guided fine-tuning strategy that first predicts possible RNA types using an RNA sequence and then appends the predicted type to the tail of sequence to refine feature embedding in a post hoc way. Our extensive evaluation across seven datasets and five tasks demonstrates the superiority of RNAErnie in both supervised and unsupervised learning. It surpasses baselines with up to 1.8% higher accuracy in classification, 2.2% greater accuracy in interaction prediction and 3.3% improved F1 score in structure prediction, showcasing its robustness and adaptability with a unified pretrained foundation.

Novel Feature Generation for Classification of Motor Activity from Functional Near-Infrared Spectroscopy Signals Using Machine Learning

Recent research in the field of cognitive motor action decoding focuses on data acquired from Functional Near-Infrared Spectroscopy (fNIRS) and its analysis. This research aims to classify two different motor activities, namely, mental drawing (MD) and spatial navigation (SN), using fNIRS data from non-motor baseline data and other motor activities. Accurate activity detection in non-stationary signals like fNIRS is challenging and requires complex feature descriptors. As a novel framework, a new feature generation by fusion of wavelet feature, Hilbert, symlet, and Hjorth parameters is proposed for improving the accuracy of the classification. This new fused feature has statistical descriptor elements, time-localization in the frequency domain, edge feature, texture features, and phase information to detect and locate the activity accurately. Three types of independent component analysis, including FastICA, Picard, and Infomax were implemented for preprocessing which removes noises and motion artifacts. Two independent binary classifiers are designed to handle the complexity of classification in which one is responsible for mental drawing (MD) detection and the other one is spatial navigation (SN). Four different types of algorithms including nearest neighbors (KNN), Linear Discriminant Analysis (LDA), light gradient-boosting machine (LGBM), and Extreme Gradient Boosting (XGBOOST) were implemented. It has been identified that the LGBM classifier gives high accuracies—98% for mental drawing and 97% for spatial navigation. Comparison with existing research proves that the proposed method gives the highest classification accuracies. Statistical validation of the proposed new feature generation by the Kruskal–Wallis H-test and Mann–Whitney U non-parametric test proves the reliability of the proposed mechanism.

An Image Analysis of River-Floating Waste Materials by Using Deep Learning Techniques

Plastic pollution in the ocean is a severe environmental problem worldwide because rivers carry plastic waste from human activities, harming the ocean’s health, ecosystems, and people. Therefore, monitoring the amount of plastic waste flowing from rivers and streams worldwide is crucial. In response to this issue of river-floating waste, our present research aimed to develop an automated waste measurement method tailored for real rivers. To achieve this, we considered three scenarios: clear visibility, partially submerged waste, and collective mass. We proposed the use of object detection and tracking techniques based on deep learning architectures, specifically the You Only Look Once (YOLOv5) and Simple Online and Realtime Tracking with a Deep Association Metric (DeepSORT). The types of waste classified in this research included cans, cartons, plastic bottles, foams, glasses, papers, and plastics in laboratory flume experiments. Our results demonstrated that the refined YOLOv5, when applied to river-floating waste images, achieved high classification accuracy, with 88% or more for the mean average precision. The floating waste tracking using DeepSORT also attained F1 scores high enough for accurate waste counting. Furthermore, we evaluated the proposed method across the three different scenarios, each achieving an 80% accuracy rate, suggesting its potential applicability in real river environments. These results strongly support the effectiveness of our proposed method, leveraging the two deep learning architectures for detecting and tracking river-floating waste with high accuracy.

Optimization of parallel SVM algorithm for big data

Parallel Support Vector Machine (SVM) based on big data has achieved some results in data mining, but due to the complexity of the data itself and a large amount of noisy data, its execution efficiency and classification accuracy in the big data environment are very low. In order to eliminate noise, a noise reduction method based on Noise Cleaning (NC) strategy was proposed, and redundant training samples in big data environments were deleted; Introduce an improved Artificial Fish Swarm Algorithm (IAFSA) to obtain the final Parallel SVM algorithm using mutual information and artificial fish swarm algorithm based on MapReduce (MIAFSA-PSVM) classification model. The results indicate that when compared to CMI-PSVM, the execution time of MIAFSA-PSVM algorithm is higher on the NDC dataset with the largest data size, The SVM parameter optimization algorithm based on MapReduce and cuckoo search (CSSVM-MR) and the particle swarm optimization based parallel support vector machine ensemble algorithm (PSO-PSVM) decreased by 40.1%, 79.3%, and 51.7%, respectively. This indicates that GIESVM-MR and MIAFSA-PSVM have strong adaptability to big data environments and high classification accuracy.

Analyzing physical activity impact based on ubiquitous nonlinear dynamics and electroencephalography data.

Understanding complex systems is made easier with the tools provided by the theory of nonlinear dynamic systems. It provides novel ideas, algorithms, and techniques for signal processing, analysis, and classification. Presently, these ideas are being applied to the investigation of how physiological signals evolve. The study applies nonlinear dynamics theory to electroencephalogram (EEG) signals to better comprehend the range of alcoholic mental states. One of the main contributions of this paper is an algorithm for automatically distinguishing between sober and drunken EEG signals based on their salient features. The study utilized various entropy-based features, including ApEn, SampEn, Shannon and Renyi entropies, PE, TS, FE, WE, and KSE, to extract information from EEG signals. To identify the most relevant features, the study employed ranking methods like T-test, Wilcoxon, and Bhattacharyya, and trained SVM classifiers with the selected features. The Bhattacharyya ranking method was found to be the most effective in achieving high classification accuracy, sensitivity, and specificity. Classification accuracy of 95.89%, the sensitivity of 94.43%, and specificity of 96.67% are achieved by the SVM classifier with radial basis function (RBF) for polynomial Kernel using the Bhattacharyya ranking method. From the result, it is clear that the model serves as a cost-effective and accurate decision-support tool for doctors in diagnosing alcoholism and for rehabilitation centres to monitor the effectiveness of interventions aimed at mitigating or reversing brain damage caused by alcoholism.

Detecting and Monitoring Artisanal Mining Operations in Semi-Arid Terrain Using Multitemporal SAR Data for InSAR Coherence Estimation and Unsupervised Classification

Abstract. The impact of artisanal mining in sub-Saharan Africa and other developing countries are often huge as they leave trails of environmental and socio-economic effects on the people and the environment. 123 Multi-temporal SAR data acquired between 2019 and 2023 was used. SARPROZ was used to co-register the images and other preprocessing steps like radiometric and geometric correction to transform the images from SAR coordinates to geographic coordinates after generating all the interferograms. A stack of the 122 coherence map was created. Unsupervised classification was implemented on the stack. Principal component analysis of dimensionality reductions method yielded a far better result than the other unsupervised cluster methods attempted, it showed a very high classification accuracy of the terrain. The principal component analysis worked by computing the covariance matrix of the stacked coherence map then performed eigen-decomposition on it to yield eigenvectors and eigenvalues, The eigenvectors corresponding to the largest eigenvalues represent the principal components. These principal components capture the directions of maximum variance, and the eigenvectors provides a reduced-dimensional representation of the image stack which is then used to reconstruct an approximation of the original image that captures the essential features of the original SAR data. Backscatter intensity of the SAR images processed for this study period for unsupervised change detection and land cover classification, delineated the different features and classes based on long term coherence values.

A battery-free, wireless, flexible bandlike e-nose based on MEMS gas sensors for precisely volatile organic compounds detection

Real-time detection of VOCs in the car cabin is crucial for protecting driver’s health. MOS gas sensors, with the advantages of low cost, simplicity and fast response, are widely used for gas monitoring. However, they typically exhibit broad-spectrum responsiveness to multiple gases. Despite numerous improvement strategies proposed from the perspective of MOS material modification, achieving ideal gas selectivity remains challenging. An e-nose based on gas sensor array is a promising approach for gas identification. However, the hardware circuit system and battery supply module in traditional e-nose result in large size, complex structure, and inconvenience in use. In this work, we synthesized MOS materials with different morphologies to form a microarray containing six MEMS gas sensors, and developed a novel belt-shaped wireless and battery-free flexible e-nose, where a mobile phone serves as its wireless power supply and data communication system. This design provides a flexible e-nose with excellent wearability and safety, allowing users to monitor surrounding gas conditions at any time conveniently. Subsequently, we design pattern recognition models using algorithms such as MLP, RF, XGBoost, and LGBM, in combination with ensemble learning strategies, achieving high accuracy in VOCs classification and low-error concentration prediction. This flexible e-nose holds significant application value for gas detection in car cabins and indoor environments.

Consistent land use and land cover classification across 20 years of various high-resolution images for detecting soil sealing in murcia, Spain

Artificial human-induced soil sealing has numerous negative consequences. The extent of impervious surfaces is a key indicator of the location and intensity of human activity; however, it is also proof of damage to the natural environment as a result of the sealing and modification of ecosystems. Remote sensing techniques can help detect and monitor changes in land use and cover over an extended period. However, the limited availability of consistent satellite images with high spatiotemporal resolutions covering several decades poses major challenges for achieving high overall classification accuracy. An accurate methodology for the multitemporal detection of artificial land cover classes was developed and applied to a case study of the metropolitan area of Murcia (Spain) with its challenging landscape conditions due to the frequent presence of bare soil. For this purpose, a variety of high-resolution satellite images from SPOT 5, Rapid Eye, and PlanetScope covering a period of 20 years were used. To improve the automated detection of built-up areas, the reflectance values of the images, normalised difference vegetation index (NDVI) and soil adjusted vegetation index (SAVI), and a building surface digital model were used as inputs for the supervised classification model. We applied a random forest algorithm to non-public, high-resolution images in the Google Earth Engine (GEE) as a processing environment to identify eight target land cover classes. The results show that the proposed methodology leads to a substantial improvement, after including the indices and the digital building model, in the overall accuracy (from 93.16 to 95.97%) and in all classes. This improvement was significant for the artificial classes and was particularly noticeable for the built-up areas (from 91.1 to 95.64%) because their confusion with bare soil was considerably reduced. This work demonstrates the effectiveness of the building-surface digital model as a tool for training the classification model, as it reduces uncertainty in confusion with other spectrally similar classes and its applicability to multisource imagery.

Bioinspired integrated triboelectric electronic tongue

An electronic tongue (E-tongue) comprises a series of sensors that simulate human perception of taste and embedded artificial intelligence (AI) for data analysis and recognition. Traditional E-tongues based on electrochemical methods suffer from a bulky size and require larger sample volumes and extra power sources, limiting their applications in in vivo medical diagnosis and analytical chemistry. Inspired by the mechanics of the human tongue, triboelectric components have been incorporated into E-tongue platforms to overcome these limitations. In this study, an integrated multichannel triboelectric bioinspired E-tongue (TBIET) device was developed on a single glass slide chip to improve the device’s taste classification accuracy by utilizing numerous sensory signals. The detection capability of the TBIET was further validated using various test samples, including representative human body, environmental, and beverage samples. The TBIET achieved a remarkably high classification accuracy. For instance, chemical solutions showed 100% identification accuracy, environmental samples reached 98.3% accuracy, and four typical teas demonstrated 97.0% accuracy. Additionally, the classification accuracy of NaCl solutions with five different concentrations reached 96.9%. The innovative TBIET exhibits a remarkable capacity to detect and analyze droplets with ultrahigh sensitivity to their electrical properties. Moreover, it offers a high degree of reliability in accurately detecting and analyzing various liquid samples within a short timeframe. The development of a self-powered portable triboelectric E-tongue prototype is a notable advancement in the field and is one that can greatly enhance the feasibility of rapid on-site detection of liquid samples in various settings.

NeuroDM: Decoding and visualizing human brain activity with EEG-guided diffusion model

Background and Objective:Brain–Computer Interface (BCI) technology has recently been advancing rapidly, bringing significant hope for improving human health and quality of life. Decoding and visualizing visually evoked electroencephalography (EEG) signals into corresponding images plays a crucial role in the practical application of BCI technology. The recent emergence of diffusion models provides a good modeling basis for this work. However, the existing diffusion models still have great challenges in generating high-quality images from EEG, due to the low signal-to-noise ratio and strong randomness of EEG signals. The purpose of this study is to address the above-mentioned challenges by proposing a framework named NeuroDM that can decode human brain responses to visual stimuli from EEG-recorded brain activity. Methods:In NeuroDM, an EEG-Visual-Transformer (EV-Transformer) is used to extract the visual-related features with high classification accuracy from EEG signals, then an EEG-Guided Diffusion Model (EG-DM) is employed to synthesize high-quality images from the EEG visual-related features. Results:We conducted experiments on two EEG datasets (one is a forty-class dataset, and the other is a four-class dataset). In the task of EEG decoding, we achieved average accuracies of 99.80% and 92.07% on two datasets, respectively. In the task of EEG visualization, the Inception Score of the images generated by NeuroDM reached 15.04 and 8.67, respectively. All the above results outperform existing methods. Conclusions:The experimental results on two EEG datasets demonstrate the effectiveness of the NeuroDM framework, achieving state-of-the-art performance in terms of classification accuracy and image quality. Furthermore, our NeuroDM exhibits strong generalization capabilities and the ability to generate diverse images.

Effect of intra-year Landsat scene availability in land cover land use classification in the conterminous United States using deep neural networks

The Landsat archive having consistent revisit times, near global extent and extensive multi-decadal temporal coverage offers a unique opportunity for land cover land use product generation. Along with this vast volume of freely available data, new classification methods based on deep learning have improved modeling capabilities. This manuscript investigates the effect of intra-annual Landsat scene availability in the accuracy of land cover land use classification in the conterminous United States. More specifically, we seek to quantify the effect of: i) increased monthly scene availability, and ii) specific months that may result in higher classification accuracy across different classes. Identifying specific months with comparable classification accuracy to the entire time series could offer significant computational gains for large-scale mapping. Our experiment incorporated deep learning classifiers and a wide range of reference data across the continental United States. Results were contrasted between five large U.S. climatic regions to further differentiate this intra-annual effect. Our findings indicate that the total number of months can have a highly variable effect in the classification accuracy ranging from minor (a few percentage points in terms of class F1 accuracy) to extremely beneficial (approaching 50% F1 improvement moving from four to twelve month observations). The benefit of increased month observations varied among climatic regions and classes: when all climate regions were combined, the grass/shrub and cultivated classes improved their F1 accuracy up to 30%, while the water class saw the least improvement of about 5%, partially due to its limited room for improvement. The effect of specific month combinations was also examined, where the total number of months was kept constant and the included months varied. The difference between the best month combination and the median combination value was estimated to be as high as about 30% for the four monthly observations scenario and the grass/shrub class. Further validation of the month selection importance comes from an example implementation scenario where F1 improvements can be as high as 10%. Our work demonstrated that month selection may offer such benefits that in some classes and climatic regions this time selection optimization is an inevitable choice due to large accuracy improvements. Also, the potential data reduction with targeted month selection would be particularly appealing to large-scale classification tasks. Due to the large extent of the climatic regions further studies are needed to quantify a more localized effect along with explanation of potential drivers.

Demonstration of In-Memory Biosignal Analysis: Novel High-Density and Low-Power 3D Flash Memory Array for Arrhythmia Detection.

Smart healthcare systems integrated with advanced deep neural networks enable real-time health monitoring, early disease detection, and personalized treatment. In this work, a novel 3D AND-type flash memory array with a rounded double channel for computing-in-memory (CIM) architecture to overcome the limitations of conventional smart healthcare systems: the necessity of high area and energy efficiency while maintaining high classification accuracy is proposed. The fabricated array, characterized by low-power operations and high scalability with double independent channels per floor, exhibits enhanced cell density and energy efficiency while effectively emulating the features of biological synapses. The CIM architecture leveraging the fabricated array achieves high classification accuracy (93.5%) for electrocardiogram signals, ensuring timely detection of potentially life-threatening arrhythmias. Incorporated with a simplified spike-timing-dependent plasticity learning rule, the CIM architecture is suitable for robust, area- and energy-efficient in-memory arrhythmia detection systems. This work effectively addresses the challenges of conventional smart healthcare systems, paving the way for a more refined healthcare paradigm.

Aquatic vegetation mapping with UAS-cameras considering phenotypes

Aquatic vegetation species at the genus level in an oxbow lake were identified in Hungary based on a multispectral Uncrewed Aerial System (UAS) survey within an elongated oxbow lake area of the Tisza River under continental climate. Seven and 13 classes were discriminated using three different classification methods (Support Vector Machine [SVM], Random Forest [RF], and Multivariate Adaptive Regression Splines [MARS]) using different input data in ten combinations: original spectral bands, spectral indices, Digital Surface Model (DSM), and Haralick texture indices. We achieved a high (97.1%) overall accuracies (OAs) by applying the SVM classifier, but the RF performed only <1% worse, as it was represented in the first places of the classification rank before the MARS. The highest classification accuracies (>84% OA) were obtained using the most important variables derived by the Recursive Feature Elimination (RFE) method. The best classification required DSM as an input variable. The poorest classification performance belonged to the model that used only texture indices or spectral indices. On the class level, Stratiotes aloides exhibit the lowest degree of separability compared to the other classes. Accordingly, we recommend using supplementary input data for the classifications besides the original spectral bands, for example, DSM, spectral, and texture indices, as these variables significantly improve the classification accuracies in the proper combinations of the input variables.

Abstract PO5-07-05: Deep learning can diagnose axillary lymph node metastases on optical virtual histologic images in breast cancer patients during surgery

Abstract Background: Reliable identification of axillary lymph node (ALN) involvement in patients with breast cancer allows for definitive axillary dissection at the time of the initial surgery, thus avoiding the need for a separate axillary surgery. However, conventional intraoperative ALN diagnostic methods are time-consuming and labor-intensive and can result in tissue destruction. Dynamic full field optical coherence tomography, also called dynamic cell imaging (DCI), has been developed and validated to offer rapid and label-free histologic approximations of metastatic and non-metastatic ALNs. In this study, we aim to optimize the diagnostic pipeline with an automated approach and present the results of using a deep learning (DL) algorithm with DCI to predict ALN status intraoperatively in patients with breast cancer. Methods: Breast cancer patients who required ALN staging were enrolled prospectively in this study. DCI was applied to bisected fresh lymph nodes in a non-destructive manner, and the specimens were subsequently sent for histopathological examination, regarded as the gold standard for comparison. A DL model was trained and fine-tuned on over 80,000 DCI images, and the results were mapped to slide level to predict ALN diagnosis. Results: Total 607 DCI slides of ALNs with 112,852 cropped patches were included in the study. The DL model was trained and validated on a dataset containing 481 slides and tested on an independent testing dataset with 126 slides. In the test set, the DL algorithm yielded accuracy for prediction of ALN status, with sensitivity and specificity of 91.9% and 95.5% and an area under the receiver operating characteristic curve (AUC) of 0.937 (95% confidence interval [CI]: 0.912-0.957) at slide level. Conclusion: These results demonstrate that the integration of DCI with DL is rapid, reduces labor requirements and minimizes tissue destruction. Meanwhile, this algorithm had high classification accuracy to predict the metastatic burden of ALNs for patients with breast cancer. Citation Format: Shuwei Zhang, Houpu Yang, Jin Zhao, Shu Wang. Deep learning can diagnose axillary lymph node metastases on optical virtual histologic images in breast cancer patients during surgery [abstract]. In: Proceedings of the 2023 San Antonio Breast Cancer Symposium; 2023 Dec 5-9; San Antonio, TX. Philadelphia (PA): AACR; Cancer Res 2024;84(9 Suppl):Abstract nr PO5-07-05.