Panagiotis Gonidakis
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Research Subjects
Panagiotis Gonidakis has a total of 56 co-authors. Here are the top co-authors by number of shared publications:
- Stefaan Poedts
KU Leuven
7 shared publications - George Miloshevich
Centre for mathematical Plasma Astrophysics (CmPA) KULeuven Leuven Belgium
5 shared publications Bart Jansen
Department of Electronics and Informatics (ETRO), Vrije Universiteit Brussel (VUB), Brussel, Belgium
3 shared publications- Ekaterina Dineva
Centre for mathematical Plasma Astrophysics, Department of Mathematics, KU Leuven, Celestijnenlaan 200B, 3001 Leuven, Belgium;
3 shared publications - Francesco Carella
Centre for mathematical Plasma Astrophysics, Department of Mathematics, KU Leuven, Celestijnenlaan 200B, 3001 Leuven, Belgium;
3 shared publications - Hyun-Jin Jeong
Centre for mathematical Plasma-Astrophysics, Department of Mathematics, KU Leuven, Celestijnenlaan 200B, 3001 Leuven, Belgium
3 shared publications
Latest Publications
- Open Access
- 11 September 2026
- Published by American Astronomical Society in The Astrophysical Journal
- Vol. 1009 (1) , 27
Statistical Analysis of Coronal Hole Spatiotemporal Evolution, Kinematics, and Lifetimes over 1.5 Solar Cycles
Coronal holes (CHs) are fundamental solar features and the primary source of high-speed solar wind, yet their long-term statistical properties remain insufficiently constrained. In this study, we leverage the Detection and Tracking Algorithm for Coronal Holes (DETACH) to construct an unprecedented 12 hr resolution CH dataset spanning 1.5 solar cycles (2010–2025) using Solar Dynamics Observatory (SDO) observations. Based on this comprehensive database, we conduct a rigorous statistical analysis of CH spatiotemporal evolution, kinematics, and lifetime dependencies. First, we reveal a prominent north–south asymmetry in various CH physical parameters, including area, magnetic flux, and magnetic field strength, providing observational constraints for hemispheric decoupling in solar dynamo models. Second, kinematic analysis demonstrates that CHs do not rotate as perfectly rigid bodies; rather, they exhibit a distinct differential rotation, albeit with a significantly smaller latitudinal gradient compared to the underlying photosphere. Furthermore, by employing survival analysis, we systematically decouple the factors influencing CH lifetimes. We find that the CH area and unsigned magnetic field strength positively correlate with lifetime, whereas extreme ultraviolet (EUV) intensity, rotation speed, and absolute latitude exhibit negative correlations with lifetime. These findings not only confirm previous theoretical and localized observational studies but also provide a robust statistical foundation for understanding large-scale CH evolution and improving space weather forecasting.
- 11 September 2026
- Published by American Astronomical Society in The Astrophysical Journal
- Vol. 1009 (1) , 27
Open AccessStatistical Analysis of Coronal Hole Spatiotemporal Evolution, Kinematics, and Lifetimes over 1.5 Solar Cycles
Coronal holes (CHs) are fundamental solar features and the primary source of high-speed solar wind, yet their long-term statistical properties remain insufficiently constrained. In this study, we leverage the Detection and Tracking Algorithm for Coronal Holes (DETACH) to construct an unprecedented 12 hr resolution CH dataset spanning 1.5 solar cycles (2010–2025) using Solar Dynamics Observatory (SDO) observations. Based on this comprehensive database, we conduct a rigorous statistical analysis of CH spatiotemporal evolution, kinematics, and lifetime dependencies. First, we reveal a prominent north–south asymmetry in various CH physical parameters, including area, magnetic flux, and magnetic field strength, providing observational constraints for hemispheric decoupling in solar dynamo models. Second, kinematic analysis demonstrates that CHs do not rotate as perfectly rigid bodies; rather, they exhibit a distinct differential rotation, albeit with a significantly smaller latitudinal gradient compared to the underlying photosphere. Furthermore, by employing survival analysis, we systematically decouple the factors influencing CH lifetimes. We find that the CH area and unsigned magnetic field strength positively correlate with lifetime, whereas extreme ultraviolet (EUV) intensity, rotation speed, and absolute latitude exhibit negative correlations with lifetime. These findings not only confirm previous theoretical and localized observational studies but also provide a robust statistical foundation for understanding large-scale CH evolution and improving space weather forecasting.
- Open Access
- 23 May 2026
- Published by American Geophysical Union (AGU) in Journal of Geophysical Research: Machine Learning and Computation
- Vol. 3 (3)
Comparing Solar Structure Detection Methods in SDO/AIA Observations and the Application to Raw Uncalibrated Data
Recent advances in solar physics increasingly rely on automated identification of coronal structures using machine learning. Yet most studies emphasize scientific performance without evaluating feasibility for onboard deployment to prioritize downlink observations. This work investigates the automated identification of active regions and coronal holes by applying segmentation and detection techniques to Solar Dynamics Observatory (SDO) data. We compare three approaches: SCSS‐Net, a deep learning model for semantic segmentation; YOLOv8n, a lightweight object detector; and a traditional pipeline based on basic computer vision operations (BCVO). Each method is assessed for its scientific accuracy and its suitability for deployment in future resource‐limited missions. While no direct hardware benchmarking is performed in this study, we assess the feasibility of onboard implementation based on the associated number of trainable parameters, architecture and hardware requirements. Training and evaluation are first conducted on well‐calibrated SDO images. We then extend the evaluation to raw and uncalibrated SDO images affected by instrumental artifacts. Performance is measured using the Intersection over Union (IoU) and Dice score. Results show that while SCSS‐Net achieves the highest segmentation quality, YOLOv8n offers a strong balance between accuracy and efficiency. The BCVO pipeline remains viable under strict hardware limitations. Interestingly, our models retain compatibility on Level‐0 observations. This is the first study comparing these widely used methods from the perspective of onboard deployment. Our findings provide a foundation for designing frameworks tailored to onboard hardware configurations.
- 23 May 2026
- Published by American Geophysical Union (AGU) in Journal of Geophysical Research: Machine Learning and Computation
- Vol. 3 (3)
Open AccessComparing Solar Structure Detection Methods in SDO/AIA Observations and the Application to Raw Uncalibrated Data
Recent advances in solar physics increasingly rely on automated identification of coronal structures using machine learning. Yet most studies emphasize scientific performance without evaluating feasibility for onboard deployment to prioritize downlink observations. This work investigates the automated identification of active regions and coronal holes by applying segmentation and detection techniques to Solar Dynamics Observatory (SDO) data. We compare three approaches: SCSS‐Net, a deep learning model for semantic segmentation; YOLOv8n, a lightweight object detector; and a traditional pipeline based on basic computer vision operations (BCVO). Each method is assessed for its scientific accuracy and its suitability for deployment in future resource‐limited missions. While no direct hardware benchmarking is performed in this study, we assess the feasibility of onboard implementation based on the associated number of trainable parameters, architecture and hardware requirements. Training and evaluation are first conducted on well‐calibrated SDO images. We then extend the evaluation to raw and uncalibrated SDO images affected by instrumental artifacts. Performance is measured using the Intersection over Union (IoU) and Dice score. Results show that while SCSS‐Net achieves the highest segmentation quality, YOLOv8n offers a strong balance between accuracy and efficiency. The BCVO pipeline remains viable under strict hardware limitations. Interestingly, our models retain compatibility on Level‐0 observations. This is the first study comparing these widely used methods from the perspective of onboard deployment. Our findings provide a foundation for designing frameworks tailored to onboard hardware configurations.
- Open Access
- 14 March 2026
- Published by Copernicus GmbH
Probabilistic Solar Flare Forecasting via Weakly Supervised Contrastive Refinement of VAE Latent Spaces
Reliable solar flare forecasting is limited by two forms of class imbalance in active region time series: (i) the overwhelming dominance of the non-flaring, quiet state over the eruptive state, and (ii) the insufficient separability between common, physically similar event classes (e.g. C-class versus M-class flares). Although empirical parameters derived from the photospheric vector magnetic field (VMF), such as those provided by SDO/HMI SHARP products, capture aspects of active region complexity and free energy buildup, they often evolve smoothly and overlap across flare classes. Consequently, while many models can distinguish between flares and no-flares reasonably well, they struggle to distinguish flare magnitude and association with eruptive phenomena (e.g. CMEs) using photospheric information alone. This suggests that improved flare-class separation requires (a) the explicit definition of what constitutes 'similarity' between pre-flare states, and (b) parametrization that emphasizes flare-relevant structure over common active region features.We investigate a representation learning strategy that combines the parametrization of SDO/HMI SHARP VMF cutouts using a Variational Autoencoder (VAE) with a contrastive stage to reshape the resulting embedding geometry. First, a VAE is trained to encode SHARP cutouts into compact latent vectors that capture active region morphology. These vectors are then refined using a Siamese-like objective constructed from weak supervision, which uses event labels and empirical SHARP parameters as proxies for elevated flare likelihood. The contrastive stage then uses this weak supervision to encourage a latent geometry that better reflects flare-relevant evolution. This study emphasizes latent-space structure, i.e. neighborhood consistency and class-conditional clustering, and evaluates whether these properties facilitate improved probabilistic prediction across multiple forecast horizons, by training lightweight downstream models on (i) empirical parameters, (ii) VAE latents and (iii) their combined representations.
- 14 March 2026
- Published by Copernicus GmbH
Open AccessProbabilistic Solar Flare Forecasting via Weakly Supervised Contrastive Refinement of VAE Latent Spaces
Reliable solar flare forecasting is limited by two forms of class imbalance in active region time series: (i) the overwhelming dominance of the non-flaring, quiet state over the eruptive state, and (ii) the insufficient separability between common, physically similar event classes (e.g. C-class versus M-class flares). Although empirical parameters derived from the photospheric vector magnetic field (VMF), such as those provided by SDO/HMI SHARP products, capture aspects of active region complexity and free energy buildup, they often evolve smoothly and overlap across flare classes. Consequently, while many models can distinguish between flares and no-flares reasonably well, they struggle to distinguish flare magnitude and association with eruptive phenomena (e.g. CMEs) using photospheric information alone. This suggests that improved flare-class separation requires (a) the explicit definition of what constitutes 'similarity' between pre-flare states, and (b) parametrization that emphasizes flare-relevant structure over common active region features.We investigate a representation learning strategy that combines the parametrization of SDO/HMI SHARP VMF cutouts using a Variational Autoencoder (VAE) with a contrastive stage to reshape the resulting embedding geometry. First, a VAE is trained to encode SHARP cutouts into compact latent vectors that capture active region morphology. These vectors are then refined using a Siamese-like objective constructed from weak supervision, which uses event labels and empirical SHARP parameters as proxies for elevated flare likelihood. The contrastive stage then uses this weak supervision to encourage a latent geometry that better reflects flare-relevant evolution. This study emphasizes latent-space structure, i.e. neighborhood consistency and class-conditional clustering, and evaluates whether these properties facilitate improved probabilistic prediction across multiple forecast horizons, by training lightweight downstream models on (i) empirical parameters, (ii) VAE latents and (iii) their combined representations.
- Open Access
- 14 March 2026
- Published by Copernicus GmbH
Machine Learning for Solar Coronal Structure Segmentation on SDO AIA Data and Applications
Automated identification of coronal structures using machine-learning techniques can support forecasting of extreme solar events, enable autonomous solar-observing missions, and accelerate understanding of physical processes in the solar atmosphere. Existing approaches typically focus on large-scale regions or adopt conservative segmentation strategies that limit structural detail. We train a lightweight variant of the You-Only-Look-Once (YOLO) object-detection framework [1] and, in parallel, design a scheme based on classical computer-vision operations and morphological filtering. Both are compared against the deep-learning-based SCSS-Net [2]. All three frameworks detect active regions and coronal holes in images from the Atmospheric Imaging Assembly onboard the Solar Dynamics Observatory. To reduce bias, training and testing use masks from multiple sources, including SPoCA [3], CHIMERA [4], Region Growth [5], and custom annotations. Methods are evaluated for scientific performance and computational cost using standard metrics such as the Dice score and Intersection over Union (IoU). We further assess on-board feasibility by outlining potential use cases and current technical limitations, and by evaluating performance on raw, uncalibrated data to ensure operational compatibility and robustness. Finally, we examine coronal hole mapping across multiple AIA wavelength channels and analyse correlations with signed and unsigned magnetic flux.References[1] Redmon et al. "You only look once: Unified, real-time object detection." Proceedings of the IEEE conference on computer vision and pattern recognition. 2016.[2] Mackovjak et al. "SCSS-Net: solar corona structures segmentation by deep learning." Monthly Notices of the Royal Astronomical Society 508.3 (2021): 3111-3124.[3] Verbeeck et al. "The SPoCA-suite: Software for extraction, characterization, and tracking of active regions and coronal holes on EUV images." Astronomy & Astrophysics 561 (2014): A29.[4] Garto et al. "Automated coronal hole identification via multi-thermal intensity segmentation." Journal of Space Weather and Space Climate 8 (2018): A02.[5] Tlatov, A., K. Tavastsherna, and V. Vasil’eva. "Coronal holes in solar cycles 21 to 23." Solar Physics 289.4 (2014): 1349-1358.
- 14 March 2026
- Published by Copernicus GmbH
Open AccessMachine Learning for Solar Coronal Structure Segmentation on SDO AIA Data and Applications
Automated identification of coronal structures using machine-learning techniques can support forecasting of extreme solar events, enable autonomous solar-observing missions, and accelerate understanding of physical processes in the solar atmosphere. Existing approaches typically focus on large-scale regions or adopt conservative segmentation strategies that limit structural detail. We train a lightweight variant of the You-Only-Look-Once (YOLO) object-detection framework [1] and, in parallel, design a scheme based on classical computer-vision operations and morphological filtering. Both are compared against the deep-learning-based SCSS-Net [2]. All three frameworks detect active regions and coronal holes in images from the Atmospheric Imaging Assembly onboard the Solar Dynamics Observatory. To reduce bias, training and testing use masks from multiple sources, including SPoCA [3], CHIMERA [4], Region Growth [5], and custom annotations. Methods are evaluated for scientific performance and computational cost using standard metrics such as the Dice score and Intersection over Union (IoU). We further assess on-board feasibility by outlining potential use cases and current technical limitations, and by evaluating performance on raw, uncalibrated data to ensure operational compatibility and robustness. Finally, we examine coronal hole mapping across multiple AIA wavelength channels and analyse correlations with signed and unsigned magnetic flux.References[1] Redmon et al. "You only look once: Unified, real-time object detection." Proceedings of the IEEE conference on computer vision and pattern recognition. 2016.[2] Mackovjak et al. "SCSS-Net: solar corona structures segmentation by deep learning." Monthly Notices of the Royal Astronomical Society 508.3 (2021): 3111-3124.[3] Verbeeck et al. "The SPoCA-suite: Software for extraction, characterization, and tracking of active regions and coronal holes on EUV images." Astronomy & Astrophysics 561 (2014): A29.[4] Garto et al. "Automated coronal hole identification via multi-thermal intensity segmentation." Journal of Space Weather and Space Climate 8 (2018): A02.[5] Tlatov, A., K. Tavastsherna, and V. Vasil’eva. "Coronal holes in solar cycles 21 to 23." Solar Physics 289.4 (2014): 1349-1358.
- Open Access
- 24 December 2025
- Published by Oxford University Press (OUP) in Monthly Notices of the Royal Astronomical Society
- Vol. 546 (1)
Classifying solar flares with k NN on sparse autoencoder representations
The high dimensionality and strong correlations of magnetic field descriptors in the Space-Weather HMI Active Region Patches (SHARPs) data set present challenges for flare prediction studies. We perform a comparison of dimensionality-reduction pipelines that combine Canonical Factor Analysis (CFA) and sparse autoencoders (SAEs), coupled with a k-Nearest Neighbours classifier, to assess how different feature representations affect predictive skill. We find that both linear (CFA) and non-linear (SAE) compression enhance flare classification relative to using the raw parameters, with the non-linear SAE yielding the highest True Skill Statistic (TSS = 0.64) for binary flare/no-flare discrimination. Multiclass predictions remain limited by the imbalance of major flares and by overlapping feature distributions, but performance is consistent across multiple configurations. Our analysis also quantifies the redundancy of SHARP parameters, showing that only 5–7 compressed features are sufficient to capture the predictive information. These results provide a benchmark for dimensionality-reduction strategies in flare forecasting and demonstrate the value of sparse representations for handling correlated astrophysical parameter spaces.
- 24 December 2025
- Published by Oxford University Press (OUP) in Monthly Notices of the Royal Astronomical Society
- Vol. 546 (1)
Open AccessClassifying solar flares with k NN on sparse autoencoder representations
The high dimensionality and strong correlations of magnetic field descriptors in the Space-Weather HMI Active Region Patches (SHARPs) data set present challenges for flare prediction studies. We perform a comparison of dimensionality-reduction pipelines that combine Canonical Factor Analysis (CFA) and sparse autoencoders (SAEs), coupled with a k-Nearest Neighbours classifier, to assess how different feature representations affect predictive skill. We find that both linear (CFA) and non-linear (SAE) compression enhance flare classification relative to using the raw parameters, with the non-linear SAE yielding the highest True Skill Statistic (TSS = 0.64) for binary flare/no-flare discrimination. Multiclass predictions remain limited by the imbalance of major flares and by overlapping feature distributions, but performance is consistent across multiple configurations. Our analysis also quantifies the redundancy of SHARP parameters, showing that only 5–7 compressed features are sufficient to capture the predictive information. These results provide a benchmark for dimensionality-reduction strategies in flare forecasting and demonstrate the value of sparse representations for handling correlated astrophysical parameter spaces.
- Open Access
- 20 August 2025
- Vol. 989 (2)
Prediction of Time-evolving Radial Magnetic Fields on the Solar Surface Using Deep Learning
- 20 August 2025
- Vol. 989 (2)
Open AccessPrediction of Time-evolving Radial Magnetic Fields on the Solar Surface Using Deep Learning
- Open Access
- 18 March 2025
- Published by Copernicus GmbH
Efficient Segmentation and Clustering of Solar Coronal Structures: A Comparison of U-Net and Classical Computer Vision Techniques Using SDO Data
Segmentation and characterization of solar coronal structures are essential for advancing our understanding of solar atmosphere and accurately identifying key regions, such as active regions and coronal holes, which are precursors to phenomena like solar flares and coronal mass ejections (CMEs). In this study, we investigate two complementary approaches to automate this process. First, we employ a previously presented deep learning-based U-Net architecture tailored for segmenting and characterizing solar coronal structures. Second, we develop a lightweight algorithm aimed at optimizing resource efficiency, consisting of classical computer vision techniques, which include thresholding and morphological filtering. The approach that best balances segmentation performance and computational efficiency will be selected for integration into a prototype designed to support future space exploration missions.To characterize the segmented regions, we propose a set of carefully designed hand-crafted features to represent and characterize the resulting segmentations. These representations are analyzed using unsupervised clustering techniques, such as K-means and t-SNE, to distinguish solar coronal structures, including active regions, coronal holes and bright points.Our dataset spans multiple layers of the solar atmosphere, incorporating HMI magnetograms (photosphere) and AIA wavelengths—94 Å (flaring regions), 171 Å (quiet Sun), 193 Å (coronal structures), and 304 Å (chromosphere). The performance of both segmentation approaches is thoroughly evaluated using metrics such as Dice score and Intersection over Union (IoU), with comparisons made against state-of-the-art methods.Future work will focus on developing feature encoding techniques to better understand and predict solar phenomena, such as solar flare emissions, while investigating the impact of different feature extraction strategies on model performance.References:Galvez, Richard, et al. "A machine-learning data set prepared from the NASA solar dynamics observatory mission." The Astrophysical Journal Supplement Series 242.1 (2019): 7. Šimon Mackovjak et al. “SCSS-Net: solar corona structures segmentation by deep learning”, Monthly Notices of the Royal Astronomical Society, Volume 508, Issue 3, December 2021, Pages 3111–3124, https://doi.org/10.1093/mnras/stab2536 Gonidakis, Panagiotis & Sóñora-Mengana, Alexander & Jansen, Bart & Vandemeulebroucke, Jef. (2023). Handcrafted Features Can Boost Performance and Data-Efficiency for Deep Detection of Lung Nodules From CT Imaging. IEEE Access. PP. 1-1. 10.1109/ACCESS.2023.3331315.
- 18 March 2025
- Published by Copernicus GmbH
Open AccessEfficient Segmentation and Clustering of Solar Coronal Structures: A Comparison of U-Net and Classical Computer Vision Techniques Using SDO Data
Segmentation and characterization of solar coronal structures are essential for advancing our understanding of solar atmosphere and accurately identifying key regions, such as active regions and coronal holes, which are precursors to phenomena like solar flares and coronal mass ejections (CMEs). In this study, we investigate two complementary approaches to automate this process. First, we employ a previously presented deep learning-based U-Net architecture tailored for segmenting and characterizing solar coronal structures. Second, we develop a lightweight algorithm aimed at optimizing resource efficiency, consisting of classical computer vision techniques, which include thresholding and morphological filtering. The approach that best balances segmentation performance and computational efficiency will be selected for integration into a prototype designed to support future space exploration missions.To characterize the segmented regions, we propose a set of carefully designed hand-crafted features to represent and characterize the resulting segmentations. These representations are analyzed using unsupervised clustering techniques, such as K-means and t-SNE, to distinguish solar coronal structures, including active regions, coronal holes and bright points.Our dataset spans multiple layers of the solar atmosphere, incorporating HMI magnetograms (photosphere) and AIA wavelengths—94 Å (flaring regions), 171 Å (quiet Sun), 193 Å (coronal structures), and 304 Å (chromosphere). The performance of both segmentation approaches is thoroughly evaluated using metrics such as Dice score and Intersection over Union (IoU), with comparisons made against state-of-the-art methods.Future work will focus on developing feature encoding techniques to better understand and predict solar phenomena, such as solar flare emissions, while investigating the impact of different feature extraction strategies on model performance.References:Galvez, Richard, et al. "A machine-learning data set prepared from the NASA solar dynamics observatory mission." The Astrophysical Journal Supplement Series 242.1 (2019): 7. Šimon Mackovjak et al. “SCSS-Net: solar corona structures segmentation by deep learning”, Monthly Notices of the Royal Astronomical Society, Volume 508, Issue 3, December 2021, Pages 3111–3124, https://doi.org/10.1093/mnras/stab2536 Gonidakis, Panagiotis & Sóñora-Mengana, Alexander & Jansen, Bart & Vandemeulebroucke, Jef. (2023). Handcrafted Features Can Boost Performance and Data-Efficiency for Deep Detection of Lung Nodules From CT Imaging. IEEE Access. PP. 1-1. 10.1109/ACCESS.2023.3331315.
- Open Access
- 15 March 2025
- Published by Copernicus GmbH
Machine Learning Algorithms for Autonomous Space Mission Operations
The Automatics in SpAce exploration (ASAP) project has as a goal the design and development of Machine Learning algorithms for the automation of operations to be implemented on the on-board processors of space missions. In the framework of ASAP a set of ML algorithms for on-board science operations of space missions have been developed/optimized on consumer-grade computing systems to be further selected for orting of existent ML models directly on an FPGA prototype. In more detail, algorithms pertaining to four main use cases have been considered: the autonomous triggering of special measurement modes and the selective downlink of plasma environment parameters; the advanced on-board data analysis of three-dimensional particle distribution functions; the on-board analysis of solar images; the on-board prediction capability of SEP related hazards. Here we describe the algorithms, their performances and requirements for the on-board implementation. ASAP has received funding from the EU’s HORIZON Research and Innovation Action (GA no.101082633)
- 15 March 2025
- Published by Copernicus GmbH
Open AccessMachine Learning Algorithms for Autonomous Space Mission Operations
The Automatics in SpAce exploration (ASAP) project has as a goal the design and development of Machine Learning algorithms for the automation of operations to be implemented on the on-board processors of space missions. In the framework of ASAP a set of ML algorithms for on-board science operations of space missions have been developed/optimized on consumer-grade computing systems to be further selected for orting of existent ML models directly on an FPGA prototype. In more detail, algorithms pertaining to four main use cases have been considered: the autonomous triggering of special measurement modes and the selective downlink of plasma environment parameters; the advanced on-board data analysis of three-dimensional particle distribution functions; the on-board analysis of solar images; the on-board prediction capability of SEP related hazards. Here we describe the algorithms, their performances and requirements for the on-board implementation. ASAP has received funding from the EU’s HORIZON Research and Innovation Action (GA no.101082633)
- Open Access
- 8 November 2023
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Access
- Vol. 11, 126221-126231
Handcrafted Features Can Boost Performance and Data-Efficiency for Deep Detection of Lung Nodules From CT Imaging
Convolutional neural networks have been widely used to detect and classify various objects and structures in computer vision and medical imaging. Access to large sets of annotated data is commonly a prerequisite for achieving good performance. Before the deep learning era, systems based on handcrafted features were employed, which typically required less annotated data but also reached inferior performance. In this work, we investigate the benefit of combining deep learning using a convolutional neural network (CNN), with handcrafted features for lung nodule detection from CT imaging. We investigate three fusion strategies with increasing complexity, and evaluate their performance for varying amounts of training data. Our results indicate that combining handcrafted features with a 3D CNN approach significantly improves lung nodule detection performance in comparison to an independently trained CNN model, regardless of the fusion strategy. Comparatively larger increases in performance were obtained when less training data was available. The fusion strategy in which features are combined with a CNN using a single end-to-end training scheme performed best overall, allowing to reduce training data by 33% to 43%, while maintaining performance. Among the investigated handcrafted features, those that describe the relative position of the candidate with respect to the lung wall and mediastinum, were found to be of most benefit.
- 8 November 2023
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Access
- Vol. 11, 126221-126231
Open AccessHandcrafted Features Can Boost Performance and Data-Efficiency for Deep Detection of Lung Nodules From CT Imaging
Convolutional neural networks have been widely used to detect and classify various objects and structures in computer vision and medical imaging. Access to large sets of annotated data is commonly a prerequisite for achieving good performance. Before the deep learning era, systems based on handcrafted features were employed, which typically required less annotated data but also reached inferior performance. In this work, we investigate the benefit of combining deep learning using a convolutional neural network (CNN), with handcrafted features for lung nodule detection from CT imaging. We investigate three fusion strategies with increasing complexity, and evaluate their performance for varying amounts of training data. Our results indicate that combining handcrafted features with a 3D CNN approach significantly improves lung nodule detection performance in comparison to an independently trained CNN model, regardless of the fusion strategy. Comparatively larger increases in performance were obtained when less training data was available. The fusion strategy in which features are combined with a CNN using a single end-to-end training scheme performed best overall, allowing to reduce training data by 33% to 43%, while maintaining performance. Among the investigated handcrafted features, those that describe the relative position of the candidate with respect to the lung wall and mediastinum, were found to be of most benefit.
- 16 March 2020
- Published by SPIE-Intl Soc Optical Eng
- Vol. 11314, 113142F
Artificially augmenting data or adding more samples? A study on a 3D CNN for lung nodule classification
Convolutional neural networks are known to require large amounts of data to achieve optimal performance. In addition, data is commonly computationally augmented using a variety of geometric and intensity transformations to further extent the set of training samples. In medical imaging, annotated data is often scarce or costly to obtain, and there is considerable interest in methods to reduce the amount of data needed. In this work, we investigate the relative benefit of increasing the amount of original data, with respect to computationally augmenting the amount of training samples, for the case of false positive reduction of lung nodules candidates. To this end, we have implemented a previously published topology for classification, shown to achieve state of the art results on the publicly available Luna16 dataset. Numerous models were trained using different amounts of unique training samples and different degrees of data augmentation involving rotations and translations, and the performance was compared. Results indicate that in general, better performance is achieved when increasing the amount of data, or augmenting the data more extensively, as expected. Surprisingly however, we observed that after reaching a certain amount of unique training samples, data augmentation leads to significantly better performance compared to adding the same number of new samples to the training dataset. We hypothesize that the augmentation has aided in learning more general {rotation and translation invariant-features, leading to improved performance on unseen data. Future experiments include more detailed characterization of this behavior, and relating this to the topology and amount of parameters to be trained.
- 16 March 2020
- Published by SPIE-Intl Soc Optical Eng
- Vol. 11314, 113142F
Artificially augmenting data or adding more samples? A study on a 3D CNN for lung nodule classification
Convolutional neural networks are known to require large amounts of data to achieve optimal performance. In addition, data is commonly computationally augmented using a variety of geometric and intensity transformations to further extent the set of training samples. In medical imaging, annotated data is often scarce or costly to obtain, and there is considerable interest in methods to reduce the amount of data needed. In this work, we investigate the relative benefit of increasing the amount of original data, with respect to computationally augmenting the amount of training samples, for the case of false positive reduction of lung nodules candidates. To this end, we have implemented a previously published topology for classification, shown to achieve state of the art results on the publicly available Luna16 dataset. Numerous models were trained using different amounts of unique training samples and different degrees of data augmentation involving rotations and translations, and the performance was compared. Results indicate that in general, better performance is achieved when increasing the amount of data, or augmenting the data more extensively, as expected. Surprisingly however, we observed that after reaching a certain amount of unique training samples, data augmentation leads to significantly better performance compared to adding the same number of new samples to the training dataset. We hypothesize that the augmentation has aided in learning more general {rotation and translation invariant-features, leading to improved performance on unseen data. Future experiments include more detailed characterization of this behavior, and relating this to the topology and amount of parameters to be trained.