Augmentation Strategies Research Articles

Automated segmentation of dental restorations using deep learning: exploring data augmentation techniques.

Deep learning has revolutionized image analysis for dentistry. Automated segmentation of dental radiographs is of great importance towards digital dentistry. The performance of deep learning models heavily relies on the quality and diversity of the training data. Data augmentation is a widely used technique implemented in machine learning and deep learning to artificially increase the size and diversity of a training dataset by applying various transformations to the original data. This work aims to automatically segment implants, prostheses, and fillings in panoramic images using 9 different deep learning segmentation models. Later, it explores the effect of data augmentation methods on segmentation performance of the models. Eight different data augmentation techniques are examined. Performance is evaluated by well-accepted metrics such as intersection over union (IoU) and Dice coefficient. While averaging the segmentation results for the three classes, IoU varies between 0.62 and 0.82 while Dice score is between 0.75 and 0.9 among deep learning models used. Augmentation techniques provided performance improvements of up to 3.37%, 5.75% and 8.75% for implant, prosthesis and filling classes, respectively. Findings reveal that choosing optimal augmentation strategies depends on both model architecture and dental structure type.

Data augmentation in predictive maintenance applicable to hydrogen combustion engines: a review

Machine-learning-based predictive maintenance models, i.e. models that predict breakdowns of machines based on condition information, have a high potential to minimize maintenance costs in industrial applications by determining the best possible time to perform maintenance. Modern machines have sensors that can collect all relevant data of the operating condition and for legacy machines which are still widely used in the industry, retrofit sensors are readily, easily and inexpensively available. With the help of this data it is possible to train such a predictive maintenance model. The main problem is that most data is obtained from normal operating conditions, whereas only limited data are from failures. This leads to highly unbalanced data sets, which makes it very difficult, if not impossible, to train a predictive maintenance model that can detect faults reliably and timely. Another issue is the lack of available real data due to privacy concerns. To address these problems, a suitable data generation strategy is needed. In this work, a literature review is conducted to identify a solution approach for a suitable data augmentation strategy that can be applied to our specific use case of hydrogen combustion engines in the automotive field. This literature review shows that, among the different state-of-the-art proposals, the most promising for the generation of reliable synthetic data are the ones based on generative models. The analysis of the different metrics used in the state of the art allows to identify the most suitable ones to evaluate the quality of generated signals. Finally, an open problem in research in this area is identified and it is the need to validate the plausibility of the data generated. The generation of results in this area will contribute decisively to the development of predictive maintenance models.

Architecture-Aware Augmentation: A Hybrid Deep Learning and Machine Learning Approach for Enhanced Parkinson’s Disease Detection

Parkinson’s Disease (PD) is a progressive neurodegenerative disorder affecting millions worldwide. Early detection is crucial for improving patient outcomes. Spiral drawing analysis has emerged as a non-invasive tool to detect early motor impairments associated with PD. This study examines the performance of hybrid deep learning and machine learning models in detecting PD using spiral drawings, with a focus on the impact of data augmentation techniques. We compare the accuracy of Vision Transformer (ViT) with K-Nearest Neighbors (KNN), Convolutional Neural Networks (CNN) with Support Vector Machines (SVM), and Residual Neural Networks (ResNet-50) with Logistic Regression, evaluating their performance on both augmented and non-augmented data. Our findings reveal that ViT with KNN, initially achieving 96.77% accuracy on unaugmented data, experienced a notable decline across all augmentation techniques, suggesting it relies heavily on global patterns in spiral drawings. In contrast, ResNet-50 with Logistic Regression showed consistent improvement with data augmentation, reaching 93.55% accuracy when rotation and flipping techniques were applied. These results highlight that hybrid models respond differently to augmentation, and careful selection of augmentation strategies is necessary for optimizing model performance. Our study provides important insights into the development of reliable diagnostic tools for early PD detection, emphasizing the need for appropriate augmentation techniques in medical image analysis.

Fog and rain augmentation for license plate recognition in tropical country environment

Automatic license plate recognition (ALPR) is a critical component in modern traffic management systems. However, ALPR systems often face challenges in accurately recognizing license plates under adverse weather conditions, such as fog and rain, prevalent in tropical regions. Deep learning ALPR models necessitate huge and diverse datasets for robustness, but data availability remains a concern since unpredictable fog and rain patterns hinder data collection. In this study, we address the issue of enhancing ALPR's robustness by introducing a novel augmentation strategy that combines traditional and weather augmentation techniques. By augmenting the dataset with weather-induced variations, we aim to improve the generalization capability of ALPR models, enabling them to handle a wider range of weather-related challenges. We also investigate the synergy between these weather augmentations and established scene text recognition (STR) methods, such as convolutional recurrent neural network (CRNN), TPS-ResNet BiLSTM-attention (TRBA), autonomous bidirectional iterative scene text recognition (ABINet), vision transformer (ViTSTR), and permutated autoregressive sequence (PARSeq), to determine their impact on recognition accuracy. Experiments using different training data sets show that training data containing a combination of traditional and weather augmentation produces the best accuracy and 1-NED performance compared to training data without augmentation and traditional augmentation only. The average increase accuracy of all STR model is 1.13% with the best increase accuracy of 3.68% using TRBA.

Fast-BEV: A Fast and Strong Bird's-Eye View Perception Baseline.

Recently, perception task based on Bird's-Eye View (BEV) representation has drawn more and more attention, and BEV representation is promising as the foundation for next-generation Autonomous Vehicle (AV) perception. However, most existing BEV solutions either require considerable resources to execute on-vehicle inference or suffer from modest performance. This paper proposes a simple yet effective framework, termed Fast-BEV, which is capable of performing faster BEV perception on the on-vehicle chips. Towards this goal, we first empirically find that the BEV representation can be sufficiently powerful without expensive transformer based transformation or depth representation. Our Fast-BEV consists of five parts, we innovatively propose (1) a lightweight deployment-friendly view transformation which fast transfers 2D image features to 3D voxel space, (2) a multi-scale image encoder which leverages multi-scale information for better performance, (3) an efficient BEV encoder which is particularly designed to speed up on-vehicle inference. We further introduce (4) a strong data augmentation strategy for both image and BEV space to avoid over-fitting, (5) a multi-frame feature fusion mechanism to leverage the temporal information. Among them, (1) and (3) enable Fast-BEV to be fast inference and deployment friendly on the on-vehicle chips, (2), (4) and (5) ensure that Fast-BEV has competitive performance. All these make Fast-BEV a solution with high performance, fast inference speed, and deployment-friendly on the on-vehicle chips of autonomous driving. Through experiments, on 2080Ti platform, our R50 model can run 52.6 FPS with 47.3% NDS on the nuScenes validation set, exceeding the 41.3 FPS and 47.5% NDS of the BEVDepth-R50 model (Li et al. 2022) and 30.2 FPS and 45.7% NDS of the BEVDet4D-R50 model (J. Huang and G. Huang, 2022). Our largest model (R101@900×1600) establishes a competitive 53.5% NDS on the nuScenes validation set. We further develop a benchmark with considerable accuracy and efficiency on current popular on-vehicle chips.

Automated Segmentation of Fetal Intracranial Volume in Three-Dimensional Ultrasound Using Deep Learning: Identifying Sex Differences in Prenatal Brain Development.

The human brain undergoes major developmental changes during pregnancy. Three-dimensional (3D) ultrasound images allow for the opportunity to investigate typical prenatal brain development on a large scale. Transabdominal ultrasound can be challenging due to the small fetal brain and its movement, as well as multiple sweeps that may not yield high-quality images, especially when brain structures are unclear. By applying the latest developments in artificial intelligence for automated image processing allowing automated training of brain anatomy in these images retrieving reliable quantitative brain measurements becomes possible at a large scale. Here, we developed a convolutional neural network (CNN) model for automated segmentation of fetal intracranial volume (ICV) from 3D ultrasound. We applied the trained model in a large longitudinal population sample from the YOUth Baby and Child cohort measured at 20- and 30-week of gestational age to investigate biological sex differences in fetal ICV as a proof-of-principle and validation for our automated method (N = 2235 individuals with 43492 ultrasounds). A total of 168 annotated, randomly selected, good quality 3D ultrasound whole-brain images were included to train a 3D CNN for automated fetal ICV segmentation. A data augmentation strategy provided physical variation to train the network. K-fold cross-validation and Bayesian optimization were used for network selection and the ensemble-based system combined multiple networks to form the final ensemble network. The final ensemble network produced consistent and high-quality segmentations of ICV (Dice Similarity Coefficient (DSC) > 0.93, Hausdorff Distance (HD): HDvoxel < 4.6 voxels, and HDphysical < 1.4 mm). In addition, we developed an automated quality control procedure to include the ultrasound scans that successfully predicted ICV from all 43492 3D ultrasounds available in all individuals, no longer requiring manual selection of the best scan for analysis. Our trained model automatically retrieved ultrasounds with brain data and estimated ICV and ICV growth in 7672 (18%) of ultrasounds in 1762 participants that passed the automatic quality control procedure. Boys had significantly larger ICV at 20-weeks (81.7 ± 0.4 mL vs. 80.8 ± 0.5 mL; B = 2.86; p = 5.7e-14) and 30-weeks (257.0 ± 0.9 mL vs. 245.1 ± 0.9 mL; B = 12.35; p = 8.2e-27) of pregnancy, and more pronounced ICV growth than girls (delta growth 0.12 mL/day; p = 1.8e-5). Our automated artificial intelligence approach provides an opportunity to investigate fetal brain development on a much larger scale and to answer fundamental questions related to prenatal brain development.

Multi-level similarity transfer and adaptive fusion data augmentation for few-shot object detection

Few-shot object detection method aims to learn novel classes through a small number of annotated novel class samples without having a catastrophic impact on previously learned knowledge, thereby expanding the trained model’s ability to detect novel classes. For existing few-shot object detection methods, there is a prominent false positive issue for the novel class samples due to the similarity in appearance features and feature distribution between the novel classes and the base classes. That is, the following two issues need to be solved: (1) How to detect these false positive samples in large-scale dataset, and (2) How to utilize the correlations between these false positive samples and other samples to improve the accuracy of the detection model. To address the first issue, an adaptive fusion data augmentation strategy is utilized to enhance the diversity of novel class samples and further alleviate the issue of false positive novel class samples. To address the second issue, a similarity transfer strategy is here proposed to effectively utilize the correlations between different categories. Experimental results demonstrate that the proposed method performs well in various settings of PASCAL VOC and MSCOCO datasets, achieving 48.7 and 11.3 on PASCAL VOC and MSCOCO under few-shot settings (shot = 1) in terms of nAP50 respectively.

A segmentation-combination data augmentation strategy and dual attention mechanism for accurate Chinese herbal medicine microscopic identification

A segmentation-combination data augmentation strategy and dual attention mechanism for accurate Chinese herbal medicine microscopic identification

Robust graph mutual-assistance convolutional networks for semi-supervised node classification tasks

Graph convolutional networks (GCNs) have made substantial progress in semi-supervised learning. However, GCNs are sensitive to graph noise, making the reconstruction of imperfect graph structures essential for adapting GCNs to complex scenarios. Prevalent GCNs employ graph augmentation to generate reliable augmented graphs, thereby mining potential graph information. Since traditional GCN layers can only perform message passing independently within each graph, GCNs primarily concentrate on fusing information from the node embeddings independently learned by the reliable augmented graphs. Nevertheless, the neglect of cross-graph information exchange during message passing limits the model's inter-graph interactions ability. In this paper, we propose a GCN framework called Robust graph mutual-assistance convolutional networks (RamNet), which introduces cross-graph information exchange mechanisms into the GCN layers. First, RamNet employs a multi-metric topology augmentation strategy to generate a synthesized graph, thereby exploring more reliable information. Second, it learns inter-graph interaction information by achieving information exchange between the raw graph and the synthesized graph during message passing. Finally, in addition to utilizing a small amount of supervised information, it also captures rich self-supervision information by maximizing consistency relationship between input graphs. We conduct comprehensive experiments in both standard and noisy scenarios, highlighting the superior performance of RamNet.

Laser weld spot detection based on YOLO-weld.

Laser weld point detection is crucial in modern industrial manufacturing, yet it faces challenges such as a limited number of samples, uneven distribution, and diverse, irregular shapes. To address these issues, this paper proposes an innovative model, YOLO-Weld, which achieves lightweight design while enhancing detection accuracy. Firstly, a targeted data augmentation strategy is employed to increase both the quantity and diversity of samples from minority classes. Following this, a Diverse Class Normalization Loss (DCNLoss)function is designed to emphasize the importance of tail data in the model's training. Secondly, the Adaptive Hierarchical Intersection over Union Loss (AHIoU Loss)function is introduced, which assigns varying levels of attention to different Intersections over Union (IoU) samples, with a particular focus on moderate IoU samples, thereby accelerating the bounding box regression process. Finally, a lightweight multi-scale feature processing module, MSBCSPELAN, is proposed to enhance multi-scale feature handling while reducing the number of model parameters. Experimental results indicate that YOLO-Weld significantly improves the accuracy and efficiency of laser weld point detection, with mean Average Precision at 50 ([Formula: see text]) and mean Average Precision at 50:95 ([Formula: see text]) increasing by 15.6% and 15.8%, respectively. Additionally, the model's parameter count is reduced by 0.4M, GFLOPS decreases by 1.1, precision improves by 4.3%, recall rises by 22.2%, and the F1 score increases by 15.1%.

The Bony Density of the Pedicle Plays a More Significant Role in the Screw Anchorage Ability Than Other Regions of the Screw Trajectory.

Osteoporosis is a crucial risk factor for screw loosening. Our studies indicate that the bone mineral density (BMD) in the screw trajectory is a better predictor of screw loosening than the BMD of the lumbar spine or the screw insertion position. Research has shown that anchorage on the screw tip is the most significant factor for screw anchorage ability, while others argue that decreased bony quality in the pedicle poses a significant risk for screw loosening. This study aimed to determine whether the bony quality of the screw tip, pedicle, or screw-anchored vertebral body plays the most significant role in screw anchorage ability. A total of 73 patients who underwent single-segment bilateral pedicle screw fixation, along with posterior and transforaminal lumbar interbody fusion (PLIF and TLIF), from March 2019 to September 2020 were included in this retrospective study. The Hounsfield unit (HU) value of the fixed vertebral bodies, the entire screw trajectory, screw tip, screw-anchoraged vertebral body, and pedicles were measured separately. Data from patients with and without screw loosening were compared, and regression analyses were conducted to identify independent risk factors. Additionally, the area under the curve (AUC) values were computed to assess the predictive performance of different parameters. Furthermore, fixation strength was calculated in numerical models with varying bony densities in different regions. HU values were found to be significantly lower in the loosening group across most measuring methods (HU values in the pedicle, 148.79 ± 97.04, 33.06 ± 34.82, p < 0.001). Specifically, the AUC of screw loosening prediction was notably higher when using HU values of the pedicle compared to other methods (AUC in the pedicle > 0.9 and in the screw insertion position > 0.7). Additionally, computational results for fixation strength revealed a clear decline in screw anchorage ability in models with poor BMD in the pedicle region. Pedicle bone quality plays a more significant role in screw anchorage ability than that in other regions. The innovation of bony augmentation strategies should pay more attention to this region to optimize the screw anchorage ability effectively.

StyleCrafter: Taming Artistic Video Diffusion with Reference-Augmented Adapter Learning

Text-to-video (T2V) models have shown remarkable capabilities in generating diverse videos. However, they struggle to produce user-desired artistic videos due to (i) text's inherent clumsiness in expressing specific styles and (ii) the generally degraded style fidelity. To address these challenges, we introduce StyleCrafter, a generic method that enhances pretrained T2V models with a style control adapter, allowing video generation in any style by feeding a reference image. Considering the scarcity of artistic video data, we propose to first train a style control adapter using style-rich image datasets, then transfer the learned stylization ability to video generation through a tailor-made finetuning paradigm. To promote content-style disentanglement, we employ carefully designed data augmentation strategies to enhance decoupled learning. Additionally, we propose a scale-adaptive fusion module to balance the influences of text-based content features and image-based style features, which helps generalization across various text and style combinations. StyleCrafter efficiently generates high-quality stylized videos that align with the content of the texts and resemble the style of the reference images. Experiments demonstrate that our approach is more flexible and efficient than existing competitors. Project page: https://gongyeliu.github.io/StyleCrafter.github.io/

Efficient generation of HPLC and FTIR data for quality assessment using time series generation model: a case study on Tibetan medicine Shilajit.

The scarcity and preciousness of plateau characteristic medicinal plants pose a significant challenge in obtaining sufficient quantities of experimental samples for quality evaluation. Insufficient sample sizes often lead to ambiguous and questionable quality assessments and suboptimal performance in pattern recognition. Shilajit, a popular Tibetan medicine, is harvested from high altitudes above 2000m, making it difficult to obtain. Additionally, the complex geographical environment results in low uniformity of Shilajit quality. To address these challenges, this study employed a deep learning model, time vector quantization variational auto- encoder (TimeVQVAE), to generate data matrices based on chromatographic and spectral for different grades of Shilajit, thereby increasing in the amount of data. Partial least squares discriminant analysis (PLS-DA) was used to identify three grades of Shilajit samples based on original, generated, and combined data. Compared with the originally generated high performance liquid chromatography (HPLC) and Fourier transform infrared spectroscopy (FTIR) data, the data generated by TimeVQVAE effectively preserved the chemical profile. In the test set, the average matrices for HPLC, FTIR, and combined data increased by 32.2%, 15.9%, and 23.0%, respectively. On the real test data, the PLS-DA model's classification accuracy initially reached a maximum of 0.7905. However, after incorporating TimeVQVAE-generated data, the accuracy significantly improved, reaching 0.9442 in the test set. Additionally, the PLS-DA model trained with the fused data showed enhanced stability. This study offers a novel and effective approach for researching medicinal materials with small sample sizes, and addresses the limitations of improving model performance through data augmentation strategies.

Pose Estimation for Cross-Domain Non-Cooperative Spacecraft Based on Spatial-Aware Keypoints Regression

Reliable pose estimation for non-cooperative spacecraft is a key technology for in-orbit service and active debris removal missions. Utilizing deep learning techniques for processing monocular camera images is effective and is a hotspot of current research. To reduce errors and improve model generalization, researchers often design multi-head loss functions or use generative models to achieve complex data augmentation, which makes the task complex and time-consuming. We propose a pyramid vision transformer spatial-aware keypoints regression network and a stereo-aware augmentation strategy to achieve robust prediction. Specifically, we primarily use the eight vertices of a cuboid satellite body as landmarks and the observable surfaces can be transformed by, respectively, using the pose labels. The experimental results on the SPEED+ dataset show that by using the existing EPNP algorithm and pseudo-label self-training method, we can achieve high-precision pose estimation for target cross-domains. Compared to other existing methods, our model and strategy are more straightforward. The entire process does not require the generation of new images, which significantly reduces the storage requirements and time costs. Combined with a Kalman filter, the robust and continuous output of the target position and attitude is verified by the SHIRT dataset. This work realizes deployment on mobile devices and provides strong technical support for the application of an automatic visual navigation system in orbit.

Use of deep learning and data augmentation by physics-based modelling for crack characterisation from multimodal ultrasonic TFM images

ABSTRACT In this article, we study a machine learning-based approach for characterising crack-like defects in specimens with complex geometries using the multi-modal total focusing method (M-TFM) applied to ultrasonic full matrix capture (FMC) data. In particular, this work frames into a scientific machine learning perspective leveraging the use of physics-based numerical solvers as an efficient yet accurate data augmentation strategy. Toward this end, the numerical simulations have been tailored to built a suitable data set aiming at gathering the uncertainties in M-TFM images encountered when the reconstruction parameters are barely known. That is, the use of such a physical knowledge has been exploited to train a deep convolutional neural network designed for retrieving crack size and position based on multi-modal TFM images. Furthermore, for validating the learning proposed scheme, an experimental campaign has been also carried out accordingly to the most relevant factors that typically impact the quality of reconstructed M-TFM images. The developed approach has been tested on both synthetic and experimental data sets never shared in the training phase.

Mixed data augmentation and osprey search strategy for enhancing YOLO in tomato disease, pest, and weed detection

Diseases, pests, and weeds are the core problem that restricts crop growth. Effective detection of biotic stress information plays an important role in crop growth and management. This paper proposes a mixed data augmentation method and the osprey search strategy for tomato biotic stress detection. Firstly, the channel attention mechanism is used to capture the channel correlation of the convolutional neural network and build an optimized object detection model YOLO_SE. Secondly, in the face of scarce and unbalanced classes, the feature transfer of “diseased leaves” based on CycleGAN and the mixed data augmentation method based on small target oversampling augmentation are used for image processing. Thirdly, a fast response osprey search strategy is proposed to optimize hyperparameters. The algorithm simulates the osprey's predatory behavior. Tent chaotic map, Lévy flight strategy, and data driven evolutionary optimization strategy are respectively used to ensure population position diversity, global optimization ability, and fast response ability. Experiments prove that the optimal validation mAP of the YOLO_SE in this paper is 2.61 % higher than that of the original model YOLOv4, and the optimal validation mAP of the expanded training set is 5.05 % higher than that of the original YOLO_SE. The optimal learning rate obtained using the osprey search strategy is 0.0011426948, and the best verification mAP of YOLO_SE trained by this learning rate is 88.42%, reaching the optimal effect.

Enhanced Semi-Supervised Medical Image Classification Based on Dynamic Sample Reweighting and Pseudo-Label Guided Contrastive Learning (DSRPGC)

In semi-supervised learning (SSL) for medical image classification, model performance is often hindered by the scarcity of labeled data and the complexity of unlabeled data. This paper proposes an enhanced SSL approach to address these challenges by effectively utilizing unlabeled data through a combination of pseudo-labeling and contrastive learning. The key contribution of our method is the introduction of a Dynamic Sample Reweighting strategy to select reliable unlabeled samples, thereby improving the model’s utilization of unlabeled data. Additionally, we incorporate multiple data augmentation strategies based on the Mean Teacher (MT) model to ensure consistent outputs across different perturbations. To better capture and integrate multi-scale features, we propose a novel feature fusion network, the Medical Multi-scale Feature Fusion Network (MedFuseNet), which enhances the model’s ability to classify complex medical images. Finally, we introduce a pseudo-label guided contrastive learning (PGC) loss function that improves intra-class compactness and inter-class separability of the model’s feature representations. Extensive experiments on three public medical image datasets demonstrate that our method outperforms existing SSL approaches, achieving 93.16% accuracy on the ISIC2018 dataset using only 20% labeled data, highlighting the potential of our approach to advance medical image classification under limited supervision.

Sample Inflation Interpolation for Consistency Regularization in Remote Sensing Change Detection

Semi-supervised learning has gained significant attention in the field of remote sensing due to its ability to effectively leverage both a limited number of labeled samples and a large quantity of unlabeled data. An effective semi-supervised learning approach utilizes unlabeled samples to enforce prediction consistency under minor perturbations, thus reducing the model’s sensitivity to noise and suppressing false positives in change-detection tasks. This principle underlies consistency regularization-based methods. However, while these methods enhance noise robustness, they also risk overlooking subtle but meaningful changes, leading to information loss and missed detections. To address this issue, we introduce a simple yet efficient method called Sample Inflation Interpolation (SII). This method leverages labeled sample pairs to mitigate the information loss caused by consistency regularization. Specifically, we propose a novel data augmentation strategy that generates additional change samples by combining existing supervised change samples with calculated proportions of change areas. This approach increases both the quantity and diversity of change samples in the training set, effectively compensating for potential information loss and reducing missed detections. Furthermore, to prevent overfitting, small perturbations are applied to the generated sample pairs and their labels. Experiments conducted on two public change detection (CD) datasets validate the effectiveness of our proposed method. Remarkably, even with only 5% of labeled training data, our method achieves performance levels that closely approach those of fully supervised learning models.

A generative-adversarial-network-based temporal raw trace data augmentation framework for fault detection in semiconductor manufacturing

In modern semiconductor manufacturing, where sophisticated process control mechanisms are standard, processing tools are equipped with sensors that generate vast amounts of raw trace data for process monitoring and fault detection. However, one of the major challenges data scientists face is the scarcity of sufficient raw trace data for defective wafers, creating an imbalance that complicates the training of machine learning models for effective fault detection. To address this issue, this paper proposes novel data augmentation structures and strategies utilizing Cycle Generative Adversarial Networks (CycleGANs) as an artificial intelligence application to synthesize temporal raw trace data for defective wafers. The effectiveness of these methods is demonstrated using a real-world dataset from the thin-film process in semiconductor fabrication. Several machine learning classification models—Gaussian Naive Bayes, Adaptive Boosting, eXtreme Gradient Boosting, and Light Gradient Boosting Machine—are employed to evaluate the performance of the augmented data. The paper identifies the optimal augmentation structure and strategy to enhance classification performance within the CycleGAN-based framework. For the thin-film processing dataset under study, the best classification performance achieves an accuracy rate of up to 99.30%, with a notably low false negative rate of 6.45%.

Data augmentation via warping transforms for modeling natural variability in the corneal endothelium enhances semi-supervised segmentation.

Image segmentation of the corneal endothelium with deep convolutional neural networks (CNN) is challenging due to the scarcity of expert-annotated data. This work proposes a data augmentation technique via warping to enhance the performance of semi-supervised training of CNNs for accurate segmentation. We use a unique augmentation process for images and masks involving keypoint extraction, Delaunay triangulation, local affine transformations, and mask refinement. This approach accurately captures the natural variability of the corneal endothelium, enriching the dataset with realistic and diverse images. The proposed method achieved an increase in the mean intersection over union (mIoU) and Dice coefficient (DC) metrics of 17.2% and 4.8% respectively, for the segmentation task in corneal endothelial images on multiple CNN architectures. Our data augmentation strategy successfully models the natural variability in corneal endothelial images, thereby enhancing the performance and generalization capabilities of semi-supervised CNNs in medical image cell segmentation tasks.