Convolutional Neural Network Model Research Articles

Machine Learning‐Based Wind Classification by Wing Deformation in Biomimetic Flapping Robots: Biomimetic Flexible Structures Improve Wind Sensing

Flying animals such as insects and birds have strain receptors on their wings. The functions and information processing capabilities of these strain receptors, however, are largely unknown. Herein, the potential ability of wing strain sensors to detect wind direction in flapping flight is experimentally explored. Seven strain gauges are attached to the biomimetic wing shafts of flexible wings. The wings flap through an electric mechanism emulating hovering flight in a wind tunnel with gentle wind, and a convolutional neural network model for wind direction classification is developed. The results show that with only one strain gauge, accurate wind classification is achieved using segmented data of ≈1 flapping cycle. Even with segmented data of only 0.2 flapping cycles, wind classification can be achieved with all seven strain gauges. Moreover, the use of wing shafts improves the classification accuracy, especially with shorter segmented data. Additional flapping phase information does not substantially improve classification performance. These results suggest that hovering animals determine the wind direction via strain sensing, which is useful for flight control. The implementation of wing strain sensors in flapping‐wing aerial robots may improve flight control ability.

Predicting malignancy in breast lesions: enhancing accuracy with fine-tuned convolutional neural network models

Predicting malignancy in breast lesions: enhancing accuracy with fine-tuned convolutional neural network models

Identification of Cactus Species Using the MobileNet Convolutional Neural Network Model

The objective of this design is to create a system capable of accurately and precisely classifying different types of succulent cacti. The system design aims to create a succulent cactus classification application using a Convolutional Neural Network (CNN) with the MobileNet architecture. The process involves collecting cactus images, dividing them into training and test datasets, and developing a CNN model to recognize patterns in succulent cactus species. The training data is used to train the CNN model, and the test data evaluates the model's accuracy. The trained model, stored in Tf.lite format, successfully classifies 15 cactus types, achieving high accuracy by employing preprocessing steps like resizing, normalization, and background removal. Over 1,200 cactus images were taken with a smartphone, categorized into 15 classes, and prepared to ensure optimal lighting, angle, background, and resolution (224x224 pixels). The MobileNet model was chosen for its high accuracy and efficiency. Hardware used includes a Samsung A54 smartphone and an Intel i7 laptop, with software such as Python, Kotlin, and Android Studio facilitating development. This design ensures the application’s accessibility, making it a valuable tool for cactus enthusiasts and the general public to easily identify different succulent cactus types. Testing of the cactus species classification program using the Convolutional Neural Network (CNN) method with the MobileNetV2 architecture yielded strong results, achieving over 90% accuracy in classifying 15 cactus species. The highest training accuracy of 0.9837 was achieved at 150 epochs without early stopping, outperforming other epoch configurations. The model successfully classified species across five main genera—Kalanchoe, Crassula, Echeveria, Haworthia, and Euphorbia. This high accuracy highlights the model's effectiveness, making it a useful tool for cactus enthusiasts and the public to accurately identify and distinguish cactus species.

An Integrative Framework for Healthcare Recommendation Systems: Leveraging the Linear Discriminant Wolf–Convolutional Neural Network (LDW-CNN) Model

In the evolving healthcare landscape, recommender systems have gained significant importance due to their role in predicting and anticipating a wide range of health-related data for both patients and healthcare professionals. These systems are crucial for delivering precise information while adhering to high standards of quality, reliability, and authentication. Objectives: The primary objective of this research is to address the challenge of class imbalance in healthcare recommendation systems. This is achieved by improving the prediction and diagnostic capabilities of these systems through a novel approach that integrates linear discriminant wolf (LDW) with convolutional neural networks (CNNs), forming the LDW-CNN model. Methods: The LDW-CNN model incorporates the grey wolf optimizer with linear discriminant analysis to enhance prediction accuracy. The model’s performance is evaluated using multi-disease datasets, covering heart, liver, and kidney diseases. Established error metrics are used to compare the effectiveness of the LDW-CNN model against conventional methods, such as CNNs and multi-level support vector machines (MSVMs). Results: The proposed LDW-CNN system demonstrates remarkable accuracy, achieving a rate of 98.1%, which surpasses existing deep learning approaches. In addition, the model improves specificity to 99.18% and sensitivity to 99.008%, outperforming traditional CNN and MSVM techniques in terms of predictive performance. Conclusions: The LDW-CNN model emerges as a robust solution for multidisciplinary disease prediction and recommendation, offering superior performance in healthcare recommender systems. Its high accuracy, alongside its improved specificity and sensitivity, positions it as a valuable tool for enhancing prediction and diagnosis across multiple disease domains.

Interpreting deep neural networks for the prediction of translation rates

BackgroundThe 5’ untranslated region of mRNA strongly impacts the rate of translation initiation. A recent convolutional neural network (CNN) model accurately quantifies the relationship between massively parallel synthetic 5’ untranslated regions (5’UTRs) and translation levels. However, the underlying biological features, which drive model predictions, remain elusive. Uncovering sequence determinants predictive of translation output may allow us to develop a more detailed understanding of translation regulation at the 5’UTR.ResultsApplying model interpretation, we extract representations of regulatory logic from CNNs trained on synthetic and human 5’UTR reporter data. We reveal a complex interplay of regulatory sequence elements, such as initiation context and upstream open reading frames (uORFs) to influence model predictions. We show that models trained on synthetic data alone do not sufficiently explain translation regulation via the 5’UTR due to differences in the frequency of regulatory motifs compared to natural 5’UTRs.ConclusionsOur study demonstrates the significance of model interpretation in understanding model behavior, properties of experimental data and ultimately mRNA translation. By combining synthetic and human 5’UTR reporter data, we develop a model (OptMRL) which better captures the characteristics of human translation regulation. This approach provides a general strategy for building more successful sequence-based models of gene regulation, as it combines global sampling of random sequences with the subspace of naturally occurring sequences. Ultimately, this will enhance our understanding of 5’UTR sequences in disease and our ability to engineer translation output.

Novel application of metabolic imaging of early embryos using a light-sheet on-a-chip device: a proof-of-concept study

Abstract STUDY QUESTION Is it feasible to safely determine metabolic imaging signatures of nicotinamide adenine dinucleotide [NAD(P)H] associated auto-fluorescence in early embryos using a light-sheet on-a-chip approach? SUMMARY ANSWER We developed an optofluidic device capable of obtaining high-resolution 3D images of the NAD(P)H autofluorescence of live mouse embryos using a light-sheet on-a-chip device as a proof-of-concept. WHAT IS KNOWN ALREADY Selecting the most suitable embryos for implantation and subsequent healthy live birth is crucial to the success rate of assisted reproduction and offspring health. Besides morphological evaluation using optical microscopy, a promising alternative is the non-invasive imaging of live embryos to establish metabolic activity performance. Indeed, in recent years, metabolic imaging has been investigated using highly advanced microscopy technologies such as fluorescence-lifetime imaging and hyperspectral microscopy. STUDY DESIGN, SIZE, DURATION The potential safety of the system was investigated by assessing the development and viability of live embryos after embryo culture for 67 h post metabolic imaging at the two-cell embryo stage (n = 115), including a control for culture conditions and sham controls (system non-illuminated). Embryo quality of developed blastocysts was assessed by immunocytochemistry to quantify trophectoderm and inner mass cells (n = 75). Furthermore, inhibition of metabolic activity (FK866 inhibitor) during embryo culture was also assessed (n = 18). PARTICIPANTS/MATERIALS, SETTING, METHODS The microstructures were fabricated following a standard UV-photolithography process integrating light-sheet fluorescence microscopy into a microfluidic system, including on-chip micro-lenses to generate a light-sheet at the centre of a microchannel. Super-ovulated F1 (CBA/C57Bl6) mice were used to produce two-cell embryos and embryo culture experiments. Blastocyst formation rates and embryo quality (immunocytochemistry) were compared between the study groups. A convolutional neural network (ResNet 34) model using metabolic images was also trained. MAIN RESULTS AND THE ROLE OF CHANCE The optofluidic device was capable of obtaining high-resolution 3D images of live mouse embryos that can be linked to their metabolic activity. The system’s design allowed continuous tracking of the embryo location, including high control displacement through the light-sheet and fast imaging of the embryos (<2 s), while keeping a low dose of light exposure (16 J · cm−2 and 8 J · cm−2). Optimum settings for keeping sample viability showed that a modest light dosage was capable of obtaining 30 times higher signal-noise-ratio images than images obtained with a confocal system (P < 0.00001; t-test). The results showed no significant differences between the control, illuminated and non-illuminated embryos (sham control) for embryo development as well as embryo quality at the blastocyst stage (P > 0.05; Yate’s chi-squared test). Additionally, embryos with inhibited metabolic activity showed a decreased blastocyst formation rate of 22.2% compared to controls, as well as a 47% reduction in metabolic activity measured by metabolic imaging (P < 0.0001; t-test). This indicates that the optofluidic device was capable of producing metabolic images of live embryos by measuring NAD(P)H autofluorescence, allowing a novel and affordable approach. The obtained metabolic images of two-cell embryos predicted blastocyst formation with an AUC of 0.974. LARGE SCALE DATA N/A. LIMITATIONS, REASONS FOR CAUTION The study was conducted using a mouse model focused on early embryo development assessing illumination at the two-cell stage. Further safety studies are required to assess the safety and use of 405 nm light at the blastocyst stage by investigating any potential negative impact on live birth rates, offspring health, aneuploidy rates, mutational load, changes in gene expression, and/or effects on epigenome stability in newborns. WIDER IMPLICATIONS OF THE FINDINGS This light-sheet on-a-chip approach is novel and after rigorous safety studies and a roadmap for technology development, potential future applications could be developed for ART. The overall cost-efficient fabrication of the device will facilitate scalability and integration into future devices if full-safety application is demonstrated. STUDY FUNDING/COMPETING INTEREST(S) This work was partially supported by an Ideas Grant (no 2004126) from the National Health and Medical Research Council (NHMRC), by the Education Program in Reproduction and Development (EPRD), Department Obstetrics and Gynaecology, Monash University, and by the Department of Mechanical and Aerospace Engineering, Faculty of Engineering, Monash University. The authors E.V-O, R.N., V.J.C., A.N., and F.H. have applied for a patent on the topic of this technology (PCT/AU2023/051132). The remaining authors have nothing to disclose.

LFL-COBC: Lightweight Federated Learning on Blockchain-Based Device Contribution Allocation

In the distributed cyber-physical systems (CPSs) within the industrial domain, the volume of data produced by interconnected devices is escalating at an unprecedented pace, presenting novel opportunities to enhance service quality through data sharing. Nevertheless, data privacy protection emerges as a significant challenge for data providers in wireless networks. This paper puts forward a solution integrating blockchain and lightweight federated learning, designated as LFL-COBC, which aims to tackle the issues related to data privacy and device performance optimization. We initially analyze multiple dimensions influencing the performance of computing devices, such as mining capacity, data quality, computational efficiency and local device deviation, which are crucial for augmenting user engagement. Based on these dimensions, we deduce a set of cooperation strategies for selecting the optimal committee members and rewarding the contributions of node devices equitably, thereby stimulating cooperation between users and servers. To intelligently and automatically detect device anomalies and alleviate the operational burden, a convolutional neural network (CNN) model is employed. Additionally, to address the escalating cost of customer participation and the potential data explosion issue, a near-optimal model pruning algorithm is designed. This algorithm can make the model obtained from the training of node equipment lightweight, thereby reducing the load of federated learning and the blockchain, as well as enhancing the overall efficiency of the system. The efficacy of our approach is demonstrated through numerical experiments on the HDFS and BGL public data sets. Experimental results indicate that the LFL-COBC scheme can effectively safeguard data privacy and optimize device performance concurrently, providing an effective solution for device anomaly detection in CPSs.

Deep learning-based Covid-19 diagnosis: a thorough assessment with a focus on generalization capabilities

The Covid-19 pandemic has significantly spurred the development of deep learning (DL) models for the pathology automatic diagnosis based on CT scan images. However, the assumption about the generalization of the proposed models remains to be assessed and shown for concrete clinical use. In this work, we have investigated the real value of widely used public datasets for the elaboration of DL models that are dedicated to automatic diagnosis of Covid-19 using CT scans. We have collected various international public datasets from 13 countries. Different Convolutional Neural Networks (CNNs) have been trained and their performances carefully assessed. Two evaluations have been conducted: (1) an internal evaluation following a cross-validation procedure, and (2) an external evaluation on real patients coming from new and different sources. The objective is to assess the generalization capabilities considering real-world conditions: different acquisition conditions, devices and configurations. Three families from the most effective CNN models have been selected (ResNet, DenseNet and EfficientNet). These have been fine-tuned, evaluated and used within a training methodology based on transfer learning. The most effective models have been further customized in order to create new models that are dedicated to the task at hand. These models have significantly improved the diagnosis performance.

Modified Multiresolution Convolutional Neural Network for Quasi-Periodic Noise Reduction in Phase Shifting Profilometry for 3D Reconstruction

Fringe profilometry is a method that obtains the 3D information of objects by projecting a pattern of fringes. The three-step technique uses only three images to acquire the 3D information from an object, and many studies have been conducted to improve this technique. However, there is a problem that is inherent to this technique, and that is the quasi-periodic noise that appears due to this technique and considerably affects the final 3D object reconstructed. Many studies have been carried out to tackle this problem to obtain a 3D object close to the original one. The application of deep learning in many areas of research presents a great opportunity to to reduce or eliminate the quasi-periodic noise that affects images. Therefore, a model of convolutional neural network along with four different patterns of frequencies projected in the three-step technique is researched in this work. The inferences produced by models trained with different frequencies are compared with the original ones both qualitatively and quantitatively.

Method of bed exit intention based on the internal pressure features in array air spring mattress

With the population ages, many patients are unable to receive comprehensive care, leading to an increase in hazardous incidents, particularly falls occurring after getting out of bed. To address this issue, this paper proposes a method for recognizing bed-exit intentions using an array air spring mattress. The method integrates convolutional neural networks with feature point matching techniques to identify both global and local features of the array air spring. For global features, a one-dimensional focal loss convolutional neural network (1D-FLCNN) model is employed to classify eight internal pressure time series and determine bed-exit status based on global features. For local features, the distribution matrix and feature point matrix of the internal pressure features are extracted to represent the spatial distribution of bed-exit postures. Euclidean distance is utilized to measure the similarity between these matrices and match bed-exit postures. Finally, the recognition results from both feature types are combined using a logical OR operation to produce the final result. Experimental validation confirms that the proposed method greatly improves the anti-interference capability and effectively avoids the problem of non-recognition due to body position and external environment.

Advancing buffet onset prediction: a deep learning approach with enhanced interpretability for aerodynamic engineering

The interaction between the shock wave and boundary layer of transonic wings can trigger periodic self-excited oscillations, resulting in transonic buffet. Buffet severely restricts the flight envelope of civil aircraft and is directly related to their aerodynamic performance and safety. Developing efficient and reliable techniques for buffet onset prediction is crucial for the advancement of civil aircraft. In this study, utilizing a comprehensive database of supercritical airfoils generated through numerical simulations, a convolutional neural network (CNN) model is firstly developed to perform buffet classification based on the flow fields. After that, employing explainable machine learning techniques, including Gradient-weighted Class Activation Mapping (Grad-CAM), random forest algorithms, and statistical analysis, the research investigates the correlations between supervised CNN features and key physical characteristics related with the separation region, shock wave, leading edge suction peak, and post-shock loading. Finally, physical buffet onset metric is established with good generalization and accuracy, providing valuable guidance for engineering design in civil aircraft.

Remote sensing image scene recognition using MobileNet

Abstract. Remote sensing image scene recognition plays a pivotal role in various applications, including environmental monitoring, disaster response, urban planning, precision agriculture, and aids in resource management and policy formulation. However, utilizing established convolutional neural networks(CNNs) models like AlexNet and VGG9 for this task can be computationally intensive and time-consuming due to their extensive parameter requirements. This dissertation introduces a MobileNet-based CNN optimized for remote sensing image scene recognition. This lightweight model significantly reduces computational load and model size without compromising accuracy, thereby enhancing efficiency. Empirical results on the NWPU45 dataset demonstrate MobileNet's superiority, achieving an accuracy of 91.16%, a Kappa coefficient of 90.96%, and an F1 score of 91.16% on the test set. Moreover, MobileNet's compact architecture, with merely 3.2531 million parameters and 587.9342 million FLOPs, underscores its efficiency and makes it a promising candidate for practical deployment in remote sensing applications. The findings suggest that MobileNet not only addresses the challenge of computational intensity but also opens new avenues for advancing scene recognition technology in the field of remote sensing.

A Review of Facial Expression Recognition Based on Convolutional Neural Networks

Abstract. The intelligent processing of human facial expressions has gained popularity due to the growth of big data and the expanding knowledge of AI and machine learning. In this paper, the author reviews the current mainstream methods for face recognition using convolutional neural network (CNN) models in deep learning, providing insights and future directions. By collecting and analyzing research findings, the characteristics, strengths and weaknesses of each model are discussed. The results indicate that CNN-based face recognition systems typically involve several steps: face detection, face alignment, face representation, face matching, and post-processing. By examining expression recognition with models like LeNet-5, AlexNet, VGGNet, and GoogleNet, the author concludes that CNNs possess features such as being data-driven, capable of feature learning, multi-level structured, adaptable to different domains, capable of real-time processing, and supporting multi-task learning. This technology has made significant strides and offers considerable potential for further development.

DEST: Difference enhanced-Swin Transformer for remote sensing change detection

ABSTRACT Remote sensing (RS) change detection (CD) has recently achieved remarkable success thanks to convolutional neural networks (CNNs). However, due to the limited receptive fields of CNN models, existing methods are prone to generating pseudo detections or missed detections. In this letter, we propose a difference enhanced-Swin Transformer network (DEST) for accurate and robust change detection in RS images. First, we design a difference enhancement module (DEM) in the feature extraction stage to boost the feature learning of differences for dual-temporal images at each level. Second, to enlarge the receptive fields of networks and capture more changed details, we apply a Swin Transformer module to the difference features to model the global contextual information. Third, to avoid the semantic loss and simultaneously solve the problem of uneven contributions of features at different levels, we design a feature weight fusion module (FWFM) to effectively aggregate multi-level feature difference maps. Extensive experimental results on two publicly available benchmarks demonstrate that the proposed method is superior to some state-of-the-art change detection models.

Efficient Brain Cancer Classification Based on Advanced CNN and Transformer Technologies

Abstract. This study addresses the challenge faced by primary medical staff in accurately distinguishing different types of brain cancer by comparing the training accuracy and iteration speed of various classification models. Specifically, thesis evaluated a basic Convolutional Neural Network (CNN) alongside two transfer learning models: EfficientNetV2 and Vision Transformer. The analysis is based on publicly available data from Kaggle, which provided a robust dataset for evaluation. The results indicate that both EfficientNetV2 and Vision Transformer models achieve significantly higher iteration convergence speeds and classification accuracy compared to the basic CNN model trained solely on the database. These advanced models demonstrate a capacity for efficient and rapid brain cancer differentiation, making them suitable for settings with limited computing resources. The study highlights that the transfer learning models can effectively support outpatient doctors by providing fast and accurate etiological screening without requiring extensive hardware. Overall, the findings suggest that EfficientNetV2 and Vision Transformer models offer a practical solution for improving brain cancer classification in primary healthcare environments, where rapid and precise diagnosis is crucial. This research underscores the potential of transfer learning techniques to enhance diagnostic capabilities while accommodating resource constraints.

An Algorithmic Study of Transformer-Based Road Scene Segmentation in Autonomous Driving

Applications such as autonomous driving require high-precision semantic image segmentation technology to identify and understand the content of each pixel in the images. Compared with traditional deep convolutional neural networks, the Transformer model is based on pure attention mechanisms, without convolutional layers or recurrent neural network layers. In this paper, we propose a new network structure called SwinLab, which is an improvement upon the Swin Transformer. Experimental results demonstrate that the improved SwinLab model achieves a segmentation accuracy comparable to that of deep convolutional neural network models in applications such as autonomous driving, with an MIoU of 77.61. Additionally, comparative experiments on the CityScapes dataset further validate the effectiveness and generalization of this structure. In conclusion, by refining the Swin Transformer, this paper simplifies the model structure, improves the training and inference speed, and maintains high accuracy, providing a more reliable semantic image segmentation solution for applications such as autonomous driving.

Cone-Beam CT to CT Image Translation Using a Transformer-Based Deep Learning Model for Prostate Cancer Adaptive Radiotherapy.

Cone-beam computed tomography (CBCT) is widely utilized in image-guided radiation therapy; however, its image quality is poor compared to planning CT (pCT), thus restricting its utility for adaptive radiotherapy (ART). Our objective was to enhance CBCT image quality utilizing a transformer-based deep learning model, SwinUNETR, which we compared with a conventional convolutional neural network (CNN) model, U-net. This retrospective study involved 260 patients undergoing prostate radiotherapy, with 245 patients used for training and 15 patients reserved as an independent hold-out test dataset. Employing a CycleGAN framework, we generated synthetic CT (sCT) images from CBCT images, employing SwinUNETR and U-net as generators. We evaluated sCT image quality and assessed its dosimetric impact for photon therapy through gamma analysis and dose-volume histogram (DVH) comparisons. The mean absolute error values for the CT numbers, calculated using all voxels within the patient's body contour and taking the pCT images as a reference, were 84.07, 73.49, and 64.69 Hounsfield units for CBCT, U-net, and SwinUNETR images, respectively. Gamma analysis revealed superior agreement between the dose on the pCT images and on the SwinUNETR-based sCT plans compared to those based on U-net. DVH parameters calculated on the SwinUNETR-based sCT deviated by < 1% from those in pCT plans. Our study showed that, compared to the U-net model, SwinUNETR could proficiently generate more precise sCT images from CBCT images, facilitating more accurate dose calculations. This study demonstrates the superiority of transformer-based models over conventional CNN-based approaches for CBCT-to-CT translation, contributing to the advancement of image synthesis techniques in ART.

Combining UAV Multispectral and Thermal Infrared Data for Maize Growth Parameter Estimation

The leaf area index (LAI) and leaf chlorophyll content (LCC) are key indicators of crop photosynthetic efficiency and nitrogen status. This study explores the integration of UAV-based multispectral (MS) and thermal infrared (TIR) data to improve the estimation of maize LAI and LCC across different growth stages, aiming to enhance nitrogen (N) management. In field trials from 2022 to 2023, UAVs captured canopy images of maize under varied water and nitrogen treatments, while the LAI and LCC were measured. Estimation models, including partial least squares regression (PLS), convolutional neural networks (CNNs), and random forest (RF), were developed using spectral, thermal, and textural data. The results showed that MS data (spectral and textural features) had strong correlations with the LAI and LCC, and CNN models yielded accurate estimates (LAI: R2 = 0.61–0.79, RMSE = 0.02–0.38; LCC: R2 = 0.63–0.78, RMSE = 2.24–0.39 μg/cm2). Thermal data reflected maize growth but had limitations in estimating the LAI and LCC. Combining MS and TIR data significantly improved the estimation accuracy, increasing R2 values for the LAI and LCC by up to 23.06% and 19.01%, respectively. Nitrogen dilution curves using estimated LAIs effectively diagnosed crop N status. Deficit irrigation reduced the N uptake, intensifying the N deficiency, while proper water and N management enhanced the LAI and LCC.

The convolutional neural network for Pacific decadal oscillation forecast

Abstract The Pacific Decadal Oscillation (PDO) is often described as a long-lived El Niño-like pattern of Pacific climate variability, and it has widespread climate and ecosystem impacts. PDO forecasts can provide useful information for policymakers on how to handle PDO impacts. Nevertheless, due to the long duration of the PDO cycles and their complex formation mechanisms, it remains a challenge to predict long lead time PDO. In this paper, we propose a transfer-learning-enhanced Convolutional Neural Network (CNN) to tackle complex ocean dynamic forecasting and predict PDO events with up to a one-year lead time. Our method first trains the CNN on historical simulations from Coupled Model Intercomparison Project 6 (CMIP6), covering the period from 1850 to 1972. This prior knowledge is then refined by further training the model with observational data from 1854 to 1972, ensuring robust performance on unseen data. Additionally, k-fold cross-validation is also employed to evaluate the model's performance across diverse subsets of data, enhancing its reliability. Throughout the testing phase from 1983 to 2022, the CNN model consistently outperforms existing dynamical forecast systems, exhibiting superior correlation skills in predicting annual mean PDO indices and PDO phases, including displaying resilience to seasonal variations. The transferred CNN is thus a powerful method to predict PDO events and is potentially valuable for a wide range of applications. This work directly supports the objectives of the World Climate Research Programme (WCRP) Grand Challenge on Climate Prediction.

Classifying driver mutations of papillary thyroid carcinoma on whole slide image: an automated workflow applying deep convolutional neural network

BackgroundInformative biomarkers play a vital role in guiding clinical decisions regarding management of cancers. We have previously demonstrated the potential of a deep convolutional neural network (CNN) for predicting cancer driver gene mutations from expert-curated histopathologic images in papillary thyroid carcinomas (PTCs). Recognizing the importance of whole slide image (WSI) analysis for clinical application, we aimed to develop an automated image preprocessing workflow that uses WSI inputs to categorize PTCs based on driver mutations.MethodsHistopathology slides from The Cancer Genome Atlas (TCGA) repository were utilized for diagnostic purposes. These slides underwent an automated tile extraction and preprocessing pipeline to ensure analysis-ready quality. Next, the extracted image tiles were utilized to train a deep learning CNN model, specifically Google’s Inception v3, for the classification of PTCs. The model was trained to distinguish between different groups based on BRAFV600E or RAS mutations.ResultsThe newly developed pipeline performed equally well as the expert-curated image classifier. The best model achieved Area Under the Curve (AUC) values of 0.86 (ranging from 0.847 to 0.872) for validation and 0.865 (ranging from 0.854 to 0.876) for the final testing subsets. Notably, it accurately predicted 90% of tumors in the validation set and 84.2% in the final testing set. Furthermore, the performance of our new classifier showed a strong correlation with the expert-curated classifier (Spearman rho = 0.726, p = 5.28 e-08), and correlated with the molecular expression-based classifier, BRS (BRAF-RAS scores) (Spearman rho = 0.418, p = 1.92e-13).ConclusionsUtilizing WSIs, we implemented an automated workflow with deep CNN model that accurately classifies driver mutations in PTCs.