Yi, Y., Zhu, D., Guo, S., Zhang, Z. & Shi, C. A review on the deterioration and approaches to enhance the durability of concrete in the marine environment. Cement Concr. Compos. 113, 103695 (2020).

Article CAS Google Scholar

Ham, Y., Han, K. K., Lin, J. J. & Golparvar-Fard, M. Visual monitoring of civil infrastructure systems via camera-equipped unmanned aerial vehicles (UAVs): A review of related works. Vis. Eng. 4(1), 18 (2016).

Article Google Scholar

Sharma, K. V. et al. Prognostic modeling of polydisperse SiO2/Aqueous glycerol nanofluids thermophysical profile using an explainable artificial intelligence (XAI) approach. Eng. Appl. Artif. Intell. 126, 106967 (2023).

Article Google Scholar

Kanti, P. K. et al. Thermophysical profile of graphene oxide and MXene hybrid nanofluids for sustainable energy applications: Model prediction with a Bayesian optimized neural network with K-cross fold validation. FlatChem 39, 100501 (2023).

Article CAS Google Scholar

Kanti, P. et al. Properties of water-based fly ash-copper hybrid nanofluid for solar energy applications: Application of RBF model. Sol. Energy Mater. Sol. Cells 234, 111423 (2022).

Article CAS Google Scholar

Kanti, P. K. et al. The stability and thermophysical properties of Al2O3-graphene oxide hybrid nanofluids for solar energy applications: application of robust autoregressive modern machine learning technique. Sol. Energy Mater. Sol. Cells 253, 112207 (2023).

Article CAS Google Scholar

Hsieh, Y. A. & Tsai, Y. J. Machine learning for crack detection: Review and model performance comparison. J. Comput. Civ. Eng. 34(5), 04020038 (2020).

Article Google Scholar

Munawar, H. S., Hammad, A. W. A., Haddad, A., Soares, C. A. P. & Waller, S. T. Image-based crack detection methods: A review. Infrastructures 6, 115. https://doi.org/10.3390/infrastructures6080115 (2021).

Article Google Scholar

Dorafshan, S., Thomas, R. J. & Maguire, M. Sdnet 2018: An annotated image dataset for non-contact concrete crack detection using deep convolutional neural networks. Data Brief 21, 16641668 (2018).

Article PubMed PubMed Central Google Scholar

C aglar, F., O zgenel, R.: Concrete crack images for classification. Mendeley Data 2 (2019)

Xu, H. et al. Automatic bridge crack detection using a convolutional neural network. Appl. Sci. 9(14), 2867 (2019).

Article Google Scholar

Harinath Reddy, C., Mini, K., Radhika, N.: Structural health monitor- ingan integrated approach for vibration analysis with wireless sensors to steel structure using image processing. In: International Conference on ISMAC in Computational Vision and Bio-Engineering, pp. 15951610 (2018). Springer

Pauly, L., Hogg, D., Fuentes, R., Peel, H.: Deeper networks for pavement crack detection. In: Proceedings of the 34th ISARC, pp. 479485 (2017). IAARC

Lins, R. G. & Givigi, S. N. Automatic crack detection and measurement based on image analysis. IEEE Trans. Instrum. Meas. 65(3), 583590. https://doi.org/10.1109/TIM.2015.2509278 (2016).

Article Google Scholar

Shahrokhinasab, E., Hosseinzadeh, N., Monirabbasi, A. & Torkaman, S. Performance of image-based crack detection systems in concrete structures. J. Soft Comput. Civ. Eng. 4(1), 127139 (2020).

Google Scholar

Munawar, H. S., Hammad, A. W., Haddad, A., Soares, C. A. P. & Waller, S. T. Image-based crack detection methods: A review. Infrastructures 6(8), 115 (2021).

Article Google Scholar

Zou, Q., Cao, Y., Li, Q., Mao, Q. & Wang, S. Cracktree: Automatic crack detection from pavement images. Pattern Recognit. Lett. 33(3), 227238 (2012).

Article Google Scholar

Salman, M., Mathavan, S., Kamal, K. & Rahman, M. Pavement crack detection using the Gabor filter. In 16th International IEEE Conference on Intelligent Transportation Systems (ITSC 2013) (eds Salman, M. et al.) 20392044 (IEEE, 2013).

Chapter Google Scholar

Niu, B., Wu, H. & Meng, Y. Application of cem algorithm in the field of tunnel crack identification. In 2020 IEEE 5th International Conference on Image, Vision and Computing (ICIVC) (eds Niu, B. et al.) 232236 (IEEE, 2020).

Chapter Google Scholar

Chhabra, G. et al. Human emotions recognition, analysis and transformation by the bioenergy field in smart grid using image processing. Electronics 11, 4059. https://doi.org/10.3390/electronics11234059 (2022).

Article Google Scholar

Baltazart, V., Nicolle, P. & Yang, L. Ongoing tests and improvements of the mps algorithm for the automatic crack detection within grey level pavement images. In 2017 25th European Signal Processing Conference (EUSIPCO) (eds Baltazart, V. et al.) 20162020 (IEEE, 2017).

Chapter Google Scholar

Jo, J. & Jadidi, Z. A high precision crack classification system using multi-layered image processing and deep belief learning. Struct. Infrastruct. Eng. 16(2), 297305 (2020).

Article Google Scholar

Landstrom, A. & Thurley, M. J. Morphology-based crack detection for steel slabs. IEEE J. Sel. Top. Signal Process. 6(7), 866875 (2012).

Article Google Scholar

Prasanna, P. et al. Automated crack detection on concrete bridges. IEEE Trans. Autom. Sci. Eng. 13(2), 591599 (2014).

Article Google Scholar

Lin, M., Zhou, R., Yan, Q. & Xu, X. Automatic pavement crack detection using hmrf-em algorithm. In 2019 International Conference on Computer, Information and Telecommunication Systems (CITS) (eds Lin, M. et al.) 15 (IEEE, 2019).

Google Scholar

Pratico, F. G., Fedele, R., Naumov, V. & Sauer, T. Detection and monitoring of bottom-up cracks in road pavement using a machine-learning approach. Algorithms 13(4), 81 (2020).

Article Google Scholar

Zhang, F. et al. A new identification method for surface cracks from uav images based on machine learning in coal mining areas. Remote Sens. 12(10), 1571 (2020).

Article Google Scholar

Zhang, L. et al. Machine learning-based real-time visible fatigue crack growth detection. Digit. Commun. Netw. 7(4), 551558 (2021).

Article Google Scholar

Dharneeshkar, J. et al. Deep learning based detection of potholes in indian roads using yolo. In 2020 International Conference on Inventive Computation Technologies (ICICT) (eds Dharneeshkar, J. et al.) 381385 (IEEE, 2020).

Google Scholar

Li, H., Zong, J., Nie, J., Wu, Z. & Han, H. Pavement crack detection algorithm based on densely connected and deeply supervised network. IEEE Access 9, 1183511842 (2021).

Article Google Scholar

Zhang, L., Yang, F., Zhang, Y. D. & Zhu, Y. J. Road crack detection using deep convolutional neural network. In 2016 IEEE International Conference on Image Processing (ICIP) (eds Zhang, L. et al.) 37083712 (IEEE, 2016).

Chapter Google Scholar

Meng, X. Concrete crack detection algorithm based on deep residual neural networks. Sci. Program. 2021, 17 (2021).

CAS Google Scholar

Su, C. & Wang, W. Concrete cracks detection using convolutional neural- network based on transfer learning. Math. Problems Eng. 2020, 110 (2020).

Google Scholar

Ye, X.-W., Jin, T. & Chen, P.-Y. Structural crack detection using deep learningbased fully convolutional networks. Adv. Struct. Eng. 22(16), 34123419 (2019).

Article Google Scholar

Feng, C. et al. Structural damage detection using deep convolutional neural network and transfer learning. KSCE J. Civ. Eng. 23(10), 44934502 (2019).

Article Google Scholar

Kim, C. N., Kawamura, K., Nakamura, H. & Tarighat, A. Automatic crack detection for concrete infrastructures using image processing and deep learning. In IOP Conference Series: Materials Science and Engineering Vol. 829 (eds Kim, C. N. et al.) 012027 (IOP Publishing, 2020).

Google Scholar

Cao, M.-T., Tran, Q.-V., Nguyen, N.-M. & Chang, K.-T. Survey on performance of deep learning models for detecting road damages using multiple dashcam image resources. Adv. Eng. Inform. 46, 101182 (2020).

Article Google Scholar

Nguyen, N. H. T., Perry, S., Bone, D., Le, H. T. & Nguyen, T. T. Two-stage convolutional neural network for road crack detection and segmentation. Expert Syst. Appl. 186, 115718 (2021).

Article Google Scholar

Park, S. E., Eem, S.-H. & Jeon, H. Concrete crack detection and quantifica- tion using deep learning and structured light. Constr. Build. Mater. 252, 119096 (2020).

Article Google Scholar

Huyan, J., Li, W., Tighe, S., Xu, Z. & Zhai, J. Cracku-net: A novel deep convolutional neural network for pixelwise pavement crack detection. Struct. Control Health Monit. 27(8), 2551 (2020).

Article Google Scholar

Kim, B., Yuvaraj, N., Sri Preethaa, K. & Arun Pandian, R. Surface crack detection using deep learning with shallow cnn architecture for enhanced computation. Neural Computing Appl. 33(15), 92899305 (2021).

Article Google Scholar

GI, K.F.: A hierarchical neural network capable of visual pattern recognition. Neural Network 1 (1989).

Russakovsky, O. et al. Imagenet large scale visual recognition challenge. Int. J. Comput. Vis. 115(3), 211252 (2015).

Article MathSciNet Google Scholar

LeCun, Y. et al. Handwritten digit recognition with a back-propagation network. Adv. Neural Inf. Process. Syst. 2, 396404 (1989).

Google Scholar

Arel, I., Rose, D. C. & Karnowski, T. P. Deep machine learning-a new frontier in artificial intelligence research [research frontier]. IEEE comput. Intel. Mag. 5(4), 1318 (2010).

Article Google Scholar

Simonyan, K., Zisserman, A.: Very deep convolutional networks for large- scale image recognition. arXiv preprint arXiv:1409.1556 (2014).

Chollet, F.: Xception: Deep learning with depthwise separable convolutions. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pp. 12511258 (2017).

He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pp. 770778 (2016).

Szegedy, C., Vanhoucke, V., Ioffe, S., Shlens, J., Wojna, Z.: Rethinking the inception architecture for computer vision. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pp. 28182826 (2016).

Szegedy, C., Ioffe, S., Vanhoucke, V., Alemi, A.A.: Inception-v4, inception-resnet and the impact of residual connections on learning. In: Thirty-first AAAI Conference on Artificial Intelligence (2017).

Andrew, G. et al. Efficient convolutional neural networks for mobile vision applications. Mobilenets. Available: http://arxiv.org/abs/1704.04861 (2017).

Sandler, M., Howard, A., Zhu, M., Zhmoginov, A., Chen, L.-C.: Mobilenetv2: Inverted residuals and linear bottlenecks. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pp. 45104520 (2018).

Huang, G., Liu, Z., Van Der Maaten, L., Weinberger, K.Q.: Densely connected convolutional networks. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pp. 47004708 (2017).

Tan, M. & Le, Q. Efficient Net: Rethinking model scaling for convolutional neural networks. In International Conference on Machine Learning (eds Tan, M. & Le, Q.) 61056114 (PMLR, 2019).

Google Scholar

Sikha, O. & Bharath, B. Vgg16-random fourier hybrid model for masked face recognition. Soft Comput. 26, 116 (2022).

Article Google Scholar

Srihari, K. & Sikha, O. Partially supervised image captioning model for urban road views. In Intelligent Data Communication Technologies and Internet of Things (eds Srihari, K. & Sikha, O.) 5973 (Springer, 2022).

Chapter Google Scholar

Originally posted here:
Transfer learned deep feature based crack detection using support vector machine: a comparative study | Scientific ... - Nature.com

From image collection to dataset creation

The whole machine learning workflow is schematically represented in Fig.1. In brief, it starts with an algorithms training which consists of three main phases, namely: (i) live-islet autofluorescence intensity imaging by exciting at 740nm and collecting in the 420460-nm range, which is dominated by NAD(P)H and lipofuscin signals; (ii) NAD(P)H auto-fluorescence lifetime imaging at the same focal plane in live islets at both low (2.2mM) and high glucose (16.7mM), with subtraction of the lipofuscin intrinsic signal, to produce metabolic data in terms of balance between free and protein-bound NAD(P)H; (iii) islet fixation and immunostaining using antibodies against glucagon and insulin to identify single and cells and then extract single-cell information from both intensity and lifetime data through spatial matching of immunofluorescence and live-islet acquisitions (Fig.1a). At this point, we curate the manual processing of experimental data to extract a set of numerical features (Fig.1b) and store them in a feature matrix. Each row of the matrix is associated with an outcome (cell identity, obtained by immunofluorescence) denoted as either or , and this is described in the target vector. At this point, the majority of the dataset is used to train a model that captures the relationship between numerical features and cell type (Fig.1c). The rest of dataset is used during the testing phase, where the performance of the model is evaluated by predicting cell type using the data portion withheld from the training phase. Upon successful completion of the testing phase, the model becomes capable of inferring cell type (i.e. the target vector) from newly collected data of sole autofluorescence and lifetime imaging, eliminating the need of performing immunostaining for cell type recognition.

General workflow. (a) Using a fluorescence microscope equipped with a FLIM (Fluorescence Lifetime Imaging Microscopy) module, data were collected from 15 Human Langerhans islets in three types of images: autofluorescence intensity (cartoon in grayscale), FLIM images, typically visualized as a phasor plot (blue cloud), and immunofluorescence images (red and green cartoon, where red represents cells, and green represents cells). (b) Single-cell data were obtained through manual segmentation of the acquired images, which resulted in one image per each segmented cell. For each cell, a number of parameters were calculated and included in a dataset then used to train a Machine Learning algorithm. (c) After a testing phase, the model can be employed to determine cell identity from new images without the need for additional immunostaining.

In more detail, to build the input dataset we performed label-free multi-photon imaging of human islets (Fig.2a), which provided two distinct types of data: islets autofluorescence intensity (Fig.2a, top panel) and lifetime (Fig.2a, center panel) micrographs. The autofluorescence signal was elicited at 740nm through multiphoton excitation and collected in the 420460-nm optical window. Each islet was measured twice: first, at 2-mM glucose concentration, which maintains a starvation condition, and then after 510min exposure to 16-mM glucose concentration, which stimulates insulin secretion from cells. Following multiphoton imaging of live islets, these were fixed and prepared for immunofluorescence (Fig.2a, bottom panel). This step involves tissue fixation, followed by permeabilization and, ultimately, incubation with anti-glucagon (red signal) and anti-insulin (green signal) antibodies to identify and cells, respectively. After image acquisition, manual segmentation (Fig.2b) was carried out to extract single-cell information: 151 features were extracted (Fig.2c) and used to construct what is referred to as the feature matrix. Each row of the matrix is associated with an outcome, specifically cell identity, denoted as either or , and this is described in the target vector. At the end, the feature matrix contains data from N=1932 cells, with each cell associated with N=151 features. In contrast, the target vector exclusively contains immunofluorescence-derived information on cell identity.

From image collection to dataset creation. (a) Human Langerhans islets' autofluorescence and lifetime are measured using label-free fluorescence microscopy, giving an autofluorescence image (top) and a phasor plot (center) of the islet as result of the live-cell imaging step. In the following step, fixation and permeabilization are performed. Then, islets are incubated with antibodies (green: anti-insulin, red: anti-glucagon), leading to the corresponding immunofluorescence image (bottom) of the islet. (b) Already obtained autofluorescence images are manually segmented by outlining Regions Of Interest (ROIs), obtaining single-cell data. Likewise the entire islet, each single cell has an associated autofluorescence image, a phasor plot, and cell identity information obtained from immunofluorescence. (c) Single-cell images are used to extract 151 features per cell, which are organized in a feature matrix. In this matrix, each row represents a single cell and each column corresponds to a specific feature. The target vector contains information about cell identity, which are derived from the immunofluorescence images.

Most of the numerical entries of the feature matrix (thecomplete list is reported in Supplementary Material)are derived from either autofluorescence intensity and lifetime data through the utilization of descriptive statistics parameters including, for instance, minimum and maximum values, trends, range of most common values, and data dispersion (Fig.3). Notably, in the optical window used for NAD(P)H detection, human islets also contain marked autofluorescence originating from lipofuscin-enriched granules20,21. These granules, byproducts of lysosomal digestion, are primarily composed of lipids and proteins, and directly correlate with age of donor19,22. Since and cells are known to possess different amounts of lipofuscin19, we decided to include a parametrization of lipofuscin granules by estimating their area normalized by the cell area. Cell morphology is instead described by three key parameters: cell area, perimeter, and circularity. Circularity quantifies how closely the cell shape resembles a perfect circle, with a value of 1 indicating a perfect circle. For what concerns autofluorescence lifetime data, the Fourier transformation converts the lifetime decay measured in each pixel of the image into a data point in the phasor plot, characterized by three parameters: the g and s coordinates, which describe the time constant of autofluorescence decay, and the frequency of observation of each specific set of g, s coordinates. Phasor clusters were quantitatively analyzed by extracting both the cluster barycenter and its standard deviation. In addition, by combining phasor-FLIM data acquired at two glucose concentrations, additional information about cell metabolism could be obtained: in fact, the shift in NAD(P)H lifetime upon glucose stimulation can be used as a descriptor of the average metabolic balance between glycolysis and oxidative phosphorylation in and cells. Finally, infrared-imaging-derived features were supplemented by adding donor-related clinical parameters (Table S1) such as age, body mass index (BMI), and the insulin stimulatory index (SI), this latter intended as the overall insulin secretion efficiency of donor-derived islets measured by a standard ELISA assay.

Overview of calculated features. In total, 151 features (in italic) have been extracted from phasor, autofluorescence, clinical, and experimental data. These features describe or summarize (in bold): Phasor plot characteristics, Cell metabolism, Cell morphology Lipofuscin content, Donor-related demographic and clinical data, Experimental conditions. In addition, various descriptive statistics parameters are used as general-purpose descriptors to summarize both autofluorescence and phasor relevant characteristics.

To facilitate the exploratory data analysis we employed the Principal Component Analysis (PCA)23 as a dimensionality reduction algorithm. We first chose the optimal number of components to avoid information loss and plot the explained variance with respect to the number of components (Fig.4a). The explained variance decreases rapidly even with few components, thus we reduced the dimensionality of the dataset from 151 to 2, making the entire dataset amenable to visualization in a 2D Cartesian plot and enabling us to observe the impact of specific features through color mapping. The PCA outcome is represented as a 19322 matrix in order to visualize only single-cell data.

Explorative data analysis with PCA and K-means clustering. The dataset dimensionality has been reduced from 151 to 2 using PCA to allow graphical representation. (a) A graphical representation of explained variance respect to the number of principal components gives an idea on how much components/dimensions are needed to retain enough information after dimensionality reduction. The explained variance drops rapidly, meaning that two components are enough to visualize the data without significant information loss. (b) The bidimensional PCA scatterplot (bottom, right) appears separated on the basis of cell type, despite mildly clustered. This suggests classification is possible using complex algorithms, maybe using a supervised approach or neural networks, as confirmed their distribution using kernel density estimation plots on the first principal component (top) and second principal component (botton, left). (c) Using experimental glucose concentration as colormap, it becomes evident that glucose concentration does not significantly affect cell classification. This implies that glucose concentration has low classification power, implying that the classification model will be able to classify cells independently of this experimental condition. (d) The elbow method allowed to choose a suitable number of clusters to have good performance by computing the WCSS (Within-Cluster Sum of Squares, i.e. sum of squared distances of all points from the centroid they belong) indicates for each iteration. The elbow (red dot) indicates the optimal number of clusters, which is 10. (e) The Gini impurity index has been computed for all clusters to assess within-cluster heterogeneity. The ideal case would be having only one class per cluster, which would result in Gini=0. However, the average Gini among all clusters is 0.37.

For instance, if data are color-coded according to cell type, and cells show mild segregation (Fig.4b, bottom right), as confirmed by kernel density estimation (KDE) plot on both the first principal component (Fig.4b, top) and second principal one (Fig.4b, bottom left), suggesting that classification might be reached, but using sophisticated supervised algorithms. If cells are color-coded by means of the glucose concentration used in the experiment (Fig.4c), it becomes challenging to accurately distinguish between and cells. This implies that glucose concentration may not possess strong classification power, thus the algorithm might be able to classify cells independently of the experimental glucose concentration used. To support this hypothesis more quantitatively the need of a Supervised Learning approach, we conducted a clustering analysis using the widely-employed k-means algorithm. First, we selected the proper amount of clusters using the elbow method. This consists in performing k-means iteratively by progressively increasing the number of clusters and calculating, for each iteration, the WCSS (Within-Cluster Sum of Squares), which represents a quantitative evaluation of how much data points are tight-bound to the cluster centroid. The optimal number of clusters should ideally match the number of classes of the classification problem (i.e. 2), but this would perform poorly here, as demonstrated by the elbow-test results (Fig.4d). The best score is reached for the highest number of clusters, but this in turn is a sign of data overfitting: the suitable number of clusters chosen was 10 (Fig.4d, red dot). For the chosen number of clusters, we assessed the performance of k-means by quantifying data heterogeneity within each cluster using the Gini impurity index (Fig.4e), exploiting the labels on the data obtained by immunofluorescence. The ideal scenario would be Gini=0, which indicates that the cluster only contains one class. Other way round, if Gini=1 (worst case), it means that data within the cluster is entirely diverse. The average Gini coefficient across all clusters is 0.37, which confirms our hypothesis about the supervised approach. To give the reader a more synthetic view of the results, we calculated the ROC_AUC (i.e. area under a ROC curve) of a two-component K-Means on PCA data, obtaining 0.60, thus reinforcing our conclusions: the explorative data analysis using PCA showed mild clustering of and cells, prompting us to use supervised classification algorithms.

Before training the model, we cleaned the dataset by manually reviewing cells, and we discarded those for which cell identity could not be confidently determined to prevent the introduction of noise into the training phase (Fig.5a). The following step involved data-preprocessing operations to favor model performance and stability: these included numerical encoding categorical features, features scaling, handling of missing values and outliers. A critical point in data preprocessing was that of addressing dataset imbalance, i.e. the unequal number of and cells in the training set. Neglecting cells from the most abundant class (i.e. cells) could lead to a biased model due to the high biological heterogeneity of Langerhans islets (Table S2)24, considering that several algorithms are built on the hypothesis of balanced classes as inputs). To address this, we employed the Synthetic Minority Oversampling Technique (SMOTE)25. This algorithm leverages existing data to generate synthetic data entries, rebalancing the : ratio of the whole dataset from 2:1 to 1:1, thus improving model training. At this point, the dataset was divided into the training and test sets (Fig. S1) to prevent overestimation of model performance during testing. Model performance and stability were further enhanced by implementing both Cross Validation and hyperparameters tuning procedures. Repeated stratified fivefold Cross Validation (with 3 repetitions) was applied, and Grid Search was chosen for cross-validation and hyperparameters tuning. The area under a ROC curve (ROC_AUC) was selected as the optimization metric, given its appropriateness for machine-learning problems based on imbalanced classes, as in this case. Four different algorithms were trained and tested (Fig.5b) using 970 cells for training (a mix of real data and synthetic data generated by SMOTE) and 216 cells (real data) for testing. Training and testing performances were then compared based on various metrics, including precision, recall, and F1 score, in addition to the area under the ROC curve. Regarding the two cell types under study (Fig.5c), cells generally exhibited scores exceeding 0.80, while cells exhibited slightly lower overall performances ranging from 0.60 to 0.70. This discrepancy may be linked to the degree of cell-type-specific information embedded in the extracted biological features. For instance, it was recently demonstrated and confirmed that cells have a significantly higher lipofuscin content compared to cells (i.e., twofold)22 and display a distinct metabolic shift toward oxidative phosphorylation upon glucose stimulation26, which is not as clearly observed in cells19. In this scenario, the extracted features convey the proper amount of information to explain the behavior of cells with confidence, while it takes more effort to take decisions on cells. All the tested algorithms showed high performance, but unsatisfactory precision or recall on -cell classification, with the exception of XGBoost. XGBoost displayed high performance and classification stability (i.e. all the computed scores were quite similar within the same class), and was thus selected for a further optimization step.

Supervised-learning results from four different models. (a) After creating the feature matrix and target vector, data undergo several preprocessing steps to enhance the performance and stability of classification. The process starts with manual cleaning, where only cells with clearly defined identities are retained in the dataset, excluding over a thousand cells, resulting in a cleaned dataset with 861 cells. Preprocessing includes encoding categorical features, handling missing values, handling outliers, and scaling the data. The dataset is then rebalanced using SMOTE (Synthetic Minority Oversampling Technique), and it is split into training and test sets. The training set, after SMOTE, comprises 970 cells and 151 features. Before training, cross-validation and hyperparameter tuning are performed to obtain a stable and high score. The model is tested on the testing data, which can be considered as new, unseen data. The original data is cleaned to improve algorithm performance. (b) Four different algorithms are tested and compared: multivariate logistic regression, boosted decision tree (XGBoost), Support Vector Machine for classification, and K-Nearest Neighbor for binary classification. Each algorithm is optimized using the most common hyperparameter range and Grid Search as the optimization algorithm. (c) Evaluation of precision, recall, F1 score, and the area under an ROC curve reveals that XGBoost is the most promising algorithm in terms of classification performance and stability. XGBoost is further optimized with Optuna, allowing for the selection of a wider hyperparameters range to improve its performance.

The optimization of XGBoost was performed by using Optuna27 that, contrary to Grid Search, does not evaluate all possible hyperparameter combinations but efficiently explores the hyperparameter space through sampling and pruning algorithms. For feature selection, we leveraged the embedded method of the XGBoost algorithm, which provides an importance score for each feature ranging from 0 to 1, based on their significance within the classification task. After an initial XGBoost training using all features, these were sorted from the most relevant to the least, and a new training phase initiated with a restricted number of features and setting different cutoff thresholds (Fig.6a). This process was aimed at enhancing model performance and, potentially, at reducing computational cost. A detailed view of all computed scores can be found in Table S3. The model with the highest performance achieved a ROC_AUC of 0.86 by using the top 116 features out of 151, thus indicating that the majority of the features are essential for optimal classification (Fig.6b). This is likely due to the high biological heterogeneity of Langerhans islet cells, both within and across donors. As mentioned earlier, Rouiller and co-workers showed that and cells disaggregated from rat islets can be separated using fluorescence-activated cell sorting (FACS). This separation relied on their intrinsic autofluorescence (mostly due to flavoproteins elicited at 488nm) and the characteristic size of the cells18. This observation prompts us to consider the significance of delving deeper into the analysis of intrinsic signals (e.g. by building a more complex algorithm as deep learning, at expense of interpretability) or by extracting more information-rich features to achieve similar or higher model performances based on standard imaging. However, a classification algorithm is needed to not underperform - or - cells classification, as evidenced by the K-Means analysis in the Explorative Data Analysis. In order to make a direct comparison with XGBoost, we applied the same pre-processing to the dataset as in the algorithm training phase, then we applied the 2-component k-Means, obtaining ROC_AUC=0.72, much lower than XGBoost (Fig.6b). Coming back to model interpretability, XGBoost has an embedded method which allows to extract and identify the most important features able to explain the classification power. By plotting the nine most important features (Fig.6c) we can observe that 6 out of 9 features are related to static autofluorescence, and the first three are able to explain more than 60% of the classification power, suggesting that most of classificatory information is encoded in the autofluorescence intensity. Indeed, by color-mapping the PCA plot (Fig. S2) for the most important feature (i.e. intensity_all_whisker_high), it can be seen that it follows the cell-type distribution shown in KDE plots. This observation is also corroborated by previous ones on the higher lipofuscin content of cells22 and their increased fluorescence intensity due to oxidative metabolism28,29,30 as compared to cells. To ensure model stability, we conducted additional assessments. First, we increased the number of folds from 5 to 10, implementing tenfold repeated stratified cross-validation. All training and Optuna-optimization steps were repeated and the same evaluation scores calculated (Table S4a), showing ROC_AUC=0.86, which is comparable to the fivefold cross-validation results (Fig.6b) together with the other metrics. Additionally, we performed the Salzberg test31, a method that involves shuffling the labels in the target vector of the training dataset, allowing the algorithm to learn from noise. This test showed a ROC_AUC=0.53, which is a 33% decrease for both training and testing (Table S4b), confirming that the model optimized during the standard training procedure was not influenced by overfitting. Furthermore, we attempted to classify data that had been excluded from training during the dataset cleaning procedure. The resulting ROC_AUC was 0.64, and all computed metrics displayed lower performance (Table S4c), thus validating the effectiveness of the cleaning procedure.

XGBoost Optimization with Optuna, Feature Importance, and Model Stability Assessment. XGBoost performances have been further improved via feature selection and larger hyperparameters tuning using Optuna. (a) We selected the optimal number of features by training XGBoost with Optuna, selecting a subset of the most important features, discovering that almost all features are needed for optimal performance. (b) The best model has been obtained for 116 features; it shows a ROC_AUC=0.86, and precision comparable with FACS on dissociated cells made by other researchers. (c) By plotting the 9 best features, we can observe that more than 50% of classification power comes almost entirely from autofluorescence images.

Go here to read the rest:
Machine-learning-guided recognition of and cells from label-free infrared micrographs of living human islets of ... - Nature.com

Find out which machine-learning stocks to get into now before the industry explodes

Source: Phonlamai Photo / Shutterstock.com

Unless you have been hiding under a rock for the past year, you are probably aware that artificial intelligence (AI) and machine-learning stocks have been leading the markets. Machine learning technology is advancing at a rapid pace. Many believe that AI will eventually replace a majority of mundane and monotonous human tasks. For corporations looking to cut costs, that is music to their ears.

Although AI has dominated the headlines recently, we are still in the early innings of what could be a trillion-dollar market in just a few years. Being an early investor in these companies can lead to exponential gains in the future. As with any emerging technology, these can be volatile stocks to hold but those patient enough should be rewarded in the long run. Here are three machine-learning stocks that could make early investors rich!

Source: Poetra.RH / Shutterstock.com

It probably comes as no surprise that we start this article off with Nvidia (NASDAQ:NVDA). Wall Street analysts have a street-high price target of $150, which implies about 15% upside from the current price.

Nvidia is the global leader in GPU chip production. These chips are used to power machine-learning infrastructure and some of the most powerful computers in the world. Nvidia just underwent a 10-for-1 stock split earlier this month. While it does not intrinsically change anything about the stock, the lower share price does make this an attractive time to load up on Nvidia stock. After a parabolic run we could see a pullback at some point but holding this stock for the next few decades should pay off in spades.

One argument from the bears against Nvidia is how expensive it is. If you compare it to its peers, Nvidia only trades at 54x forward earnings which is cheaper than both Advanced Micro Devices (NASDAQ:AMD) and Broadcom (NASDAQ:AVGO). If there is one machine-learning stock to buy early and own forever, it is Nvidia.

Source: Sundry Photography / Shutterstock.com

Palantir Technologies (NYSE:PLTR) has been a controversial stock since it went public in 2020. Dont believe us? The current one-year price target range from Wall Street analysts ranges between $9 to $35. Palantir is trading slightly above its average price target of $21.45 after gaining more than 40% so far in 2024.

The main question that always comes up with Palantir is about what the company does as a business. Palantir sells data analytics software platforms to enterprises and governments around the world. This platform can instantly organize and analyze large sets of data while using machine learning to complete these tasks more efficiently for the user. In late 2023, Palantir was named the worlds top vendor for AI and machine learning software.

One problem with Palantirs stock is how expensive it trades compared to the companys growth rate. Shares trade at 19.4x forward sales and 71x forward earnings. These multiples put Palantir well ahead of other similarly sized software. You will often pay a premium for industry leaders and Palantir is on track to being the top machine-learning software provider in the world.

Source: rblfmr / Shutterstock.com

Snowflake (NYSE:SNOW) is another stock that has had a controversial start to its public tenure. The stock debuted on Wall Street around the same time that Palantir did but the two stocks have gone in opposite directions since then. Analysts remain bullish on the stock with an average price target of $209.97 and a street-high target of $600!

Throughout the years, Snowflake and Palantir have often been compared to each other. Both companies produce software platforms that provide data analytics for their users. Snowflake offers plenty of different machine-learning tools including its Snowpark ML platform. The company has partnered with Nvidia to provide users with generative AI in the data cloud. Snowflakes enterprise-grade large language models (LLM) known as Arctic also use Nvidias powerful AI software to provide optimized performance.

All of that is to say that Snowflake is emerging as a machine-learning powerhouse at a time when its stock has struggled. It is trading at historically low multiples, including at 13.9x sales compared to its five-year average multiple of 42.9x. Snowflake is also trading at 47x free cash flow compared to its five-year average of 906x. The stock still isnt cheap, but years from now this dip will be looked at as the opportunity of a lifetime.

On the date of publication, Ian Hartana and Vayun Chugh did not hold (either directly or indirectly) any positions in the securities mentioned in this article. The opinions expressed in this article are those of the writer, subject to the InvestorPlace.comPublishing Guidelines.

Chandler Capital is the work of Ian Hartana and Vayun Chugh. Ian Hartana and Vayun Chugh are both self-taught investors whose work has been featured in Seeking Alpha. Their research primarily revolves around GARP stocks with a long-term investment perspective encompassing diverse sectors such as technology, energy, and healthcare.

Original post:
Rags to Riches: 3 Machine Learning Stocks That Could Make Early Investors Rich - InvestorPlace

QC Ware to bring Q2B conferenceto Tokyo & Asia for a third time, connecting the global and Asian quantum computing ecosystems and bringing together quantum industry experts across computing, telecommunications, sensing, finance, automotive and more.

PALO ALTO, Calif. and TOKYO, June 24, 2024 /PRNewswire/ -- QC Ware, a leading provider of industry-disrupting quantum and quantum-inspired machine learning and chemistry simulation solutions, today announced the third edition of its Q2B Tokyo Conference, to take place on July 24-25, 2024.

The conference, which will be held at the Grand Hyatt Tokyo, will dive deep into all major quantum technologies and themes: computing, sensing, communications/security, error correction, quantum AI, HPC integration and more. Attendees can expect to see featured keynotes, industry case studies and discussions led by experts at the forefront of quantum R&D from some of the world's leading businesses and institutions across government, academia, and Fortune 100 companies.

"Our team is always super excited to be back in Tokyo for Q2B, especially given the strength of the Japanese quantum ecosystem and its willingness to attract the attention of an already captivated international community," saidQC Ware CEO Matt Johnson. "The quantum computing ecosystem in Asia, and specifically in Japan, is incredibly vibrant and exciting, and the work undertaken to directly and indirectly create that environment cannot be overstated."

Through keynotes, business seminars, breakout sessions, technical workshops, and panel discussions, attendees at Q2B Tokyo will be able to learn about the latest hardware breakthroughs and applications in optimization, chemistry simulations, pharmaceutical and materials discovery, error correction and machine learning. Additionally, the conference will feature several panels and sessions from real-world practitioners, end users and experts across industries. Notable speakers include:

Q2B is dedicated to promoting and growing the quantum technologies ecosystem. One exciting component of the conference is the start-up track, which includes a pitch competition and a "New Faces" panel. The pitch competition, organized by QAI Ventures, will offer the winner a golden ticket to the QAI Ventures speed dating event. This can lead to participation in the early-stage quantum tech startup QAI Accelerator and an investment up to CHF 200k. The New Faces panel will showcase new and exciting companies in their Q2B debut.

Finally, leaders charting the future of quantum technology adoption will deliver sessions on recent advancements in the field, including:

There will be plenty to see on the exhibit floor with a number of vendors showcasing their latest advancements in quantum technologies, including: Keysight, Quantinuum, QuEra Computing, Quemix, PsiQuantum, Q-CTRL, Strangeworks, Qedma, Classiq, C12, Pasqal, Oxford Quantum Circuits, AWS, Equal1, Norma, Carrousel Digital, and more.

Register to attend Q2B 2024 Tokyo here.

About QC Ware

QC Wareis a leading software and services company at the forefront of quantum and classical computing. Our team includes some of the industry's top experts in machine learning and chemistry applications for near-term quantum computers and high-performance classical systems. We deliver real enterprise value with cutting-edge computational technology, providing innovative solutions that drive business success. QC Ware is headquartered in Palo Alto, California, and supports its European customers through its subsidiary in Paris and customers in Asia through its business development office in Tokyo, Japan. QC Ware also organizes Q2B, a global series of conferences for industry, practitioner, and academic quantum computing communities.

Media Contact:[emailprotected]

Logo - https://mma.prnewswire.com/media/2343975/4777000/Q2B24_Paris_Logo.jpg

Read this article:
QC Ware Announces 3rd Q2B Tokyo Conference, Focusing on the Roadmap to Quantum Value in Asia and Beyond - PR Newswire

The hype around AI is pervasive, and organizations across industries are investing in what has quickly become a must-have technology. As fast as this new AI arms race has heated up, many organizations are left wondering if they are already late, and that fear of missing out (FOMO) can quickly turn into a panic that they are behind the competition and destined to be on the outside looking in.

Spoiler alert: they are not.

However, its not because this is a case of unwarranted hypeadoption of AI technology is growing rapidlybut the sense of get on board with everything now because the last train is leaving the station is overplayed. Furthermore, FOMO is never a good reason to do somethingespecially deploying a cutting-edge technology that is still evolving with new use cases being thought up and brought to market at light speed. Per a recent study, 42% of CIOs do not expect to see positive ROI in their AI investments for at least 2-3 years. The significant time and resource investment in AI is a marathonnot a sprint. So, then, what is an organization to do? Maintain focus, find your center, and execute. Here is how.

First, we must acknowledge that we have seen a version of this movie before. The market gets a disruptive new technology (PC, Dotcom, Smartphone, Big Data, SaaS, Cloud, etc.) that promises to fundamentally shift the way business is done, and organizations immediately react as though everything must go to it right away. Let us look at one of these examples. With the cloud, use cases were wide open, and organizations jumped in headlong, ceding massive amounts of data, control, and even large budgets to a handful of companies with no plan B.

Only later, though, did they realize that these benefits came with some tangible challenges: vendor lock impacting the promise of cheap scalability; concerns about visibility, accessibility, and security of data; ability to integrate with other areas of their organizations; among others. While the wonderful promise of cloud has come to fruition, the industry has also learned that there is no one-size-fits-all approach, and organizations must be deliberate in their strategies.

The key lesson learned from the disruption brought on by cloud technologies was the importance of maintaining focus on the desired outcome before blindly allocating IT resources. Today, the same holds true with AI. As its usage becomes increasingly pervasive, organizations must strike the right balance between exploring its wide-open adoption and concentrating investment on clearly defined projects that will produce tangible and measurable results.

There is a sense of urgency in all organizations wanting to deploy AI, but they must also remain grounded in best practices and lessons learned. In short, yes. They need to get on board, but they need to be realistic and deploy toward a tangible result where ROI is noticeable and impactful not just a sweeping promise.

One of the more recent challenges with determining the correct course of action with AI is dierentiating among the variety of AI/ML technologies. For example, while most organizations understand the main dierences between traditional AI andGenerative AI, they also exist in a host of dierent flavors and come with a constant stream of new concepts and models. The technology itself can be completely overwhelming, and per the recent CIO survey, less than half (49%) believe their IT departments have AI ready technical skills. Confidence in AI readiness wanes even further when considering other areas/functions within the organization such as data and analytics reporting or security infrastructure.

Yet, flashy and exciting as it is, the technology is only part of itthe what. Organizations need to maintain their focus on the why. It is important to take time to understand the models available, but it is critical to tie them to use cases and outcomes when making decisions about which make sense to implement. Organizations should know their own pain points and areas for improvement, and they can always partner with industry leaders who can help them determine what AI technology to use and where AI can have an immediate positive impact.

Organizations should get started in an area where they will see resultswhatever those may be (new revenue streams, call center productivity, impact on employee experience, revamps of painful processes, etc.). Furthermore, the results do not immediately need to have a massive impact on the bottom line to count. They simply need to be positive, tangible, and measurable. Success builds on itself. From there, organizations can build on that early momentum to find more and more use cases that would benefit from AI, but they must still maintain their focus on their own desired outcomes. This requires blocking out the noise and market FUD about competitors or what some aspirational companies are doing with AI.

Market FUD is especially dangerous when organizations are comparing themselves to their industry peers, as it can be easy to look at competitors and see early examples of their AI successes. However, in many cases, these ideas or concepts are the ones they want to share, and organizations are far from the promise of scaling or even implementing AI across their businesses.

With the unrelenting hype around AI, it is easy to see how organizations can feel that they need to go all in on the technology right away for fear of being left behind. The reality, though, is that organizations are not launching AI initiatives at any kind of scalethat will take time.

The smartest ones are not standing still and are finding the specific areas in which they can apply AI to drive a tangible positive result and build on these early successes. Furthermore, they are not navigating this solo. The endless stream of innovations and models are a good reminder to seek the expertise that can help make an AI investment go smoothly. As with every disruptive technology, aligning with the right partner who understands AI from all angles can help make informed decisions on how to go about getting started. From an initial project showing positive results, organizations can set the stage for additional use cases and begin to scale their AI eorts.

[To share your insights with us, please write topsen@itechseries.com]

Read more:
Don't Panic: Why AI FOMO is Overblown - AiThority