Most valuable datasets in Radiology and Nuclear Medicine are small. This can be seen e.g. by the amount of subjects/patients in radiomics studies, where most studies go from 20 to 100 patients. Radiomics studies apply machine learning to the extracted radiomics features and this methodology can be used when the dataset size is not appropriate for Deep Learning. However, Deep Learning has consistently showed superior results when sufficient data is available. The main issue is that large amounts of labelled data are usually not available and the reasons behind are data unavailability (low prevalence diseases), data protection and data curation process.
This Research Topic will present and address solutions to work with small datasets and possible solutions to tackle all related issues. Transfer Learning has been considered one of the best techniques to work with small datasets. It is based on the idea that you can transfer knowledge from one task to another. However, more deep learning techniques can be used such as Bayesian Deep Learning or synthetic data generation as data augmentation. Alternatively, solutions to increase the dataset such as better annotation tools, possibly including Active Learning techniques or Federated Learning are also important in the context of small datasets.
Issues such as data heterogeneity and fairness are also of interest. Data heterogeneity (from different scanners and protocols used for imaging) is particularly important in the case of small datasets and to handle it, several approaches can harmonize the data or train the algorithm in order to improve performance. Fairness is fundamental for the future of Artificial Intelligence and awareness of biases when developing models needs to increase. Solutions to mitigate such biases are also of extreme relevance.
Working with small datasets is still difficult, but possible solutions have been presented for fields outside of Radiology and Nuclear Medicine. A perspective of possible solutions in Radiology and Nuclear Medicine, comparisons, and new out-of-box solutions are sought in this Research Topic. Analysis with respect to the number of patients as well as datasets with heterogeneous or imbalanced datasets is welcome in this Research Topic. Furthermore, analysis of biases and fairness in the context of small datasets is highly appreciated.
We are particularly interested in the articles that cover the following themes:
• Federated Learning
• Transfer Learning
• Bayesian Deep Learning
• Regularisation techniques for Deep Learning
• Synthetic Data Generation
• Unsupervised Learning
• Active Learning
• Fairness in data and models
• Data Heterogeneity
• Comparison of Radiomics and Deep Learning strategies
• Comparison of different Deep Learning strategies
Keywords:
small datasets, artificial intelligence
Important Note:
All contributions to this Research Topic must be within the scope of the section and journal to which they are submitted, as defined in their mission statements. Frontiers reserves the right to guide an out-of-scope manuscript to a more suitable section or journal at any stage of peer review.
Most valuable datasets in Radiology and Nuclear Medicine are small. This can be seen e.g. by the amount of subjects/patients in radiomics studies, where most studies go from 20 to 100 patients. Radiomics studies apply machine learning to the extracted radiomics features and this methodology can be used when the dataset size is not appropriate for Deep Learning. However, Deep Learning has consistently showed superior results when sufficient data is available. The main issue is that large amounts of labelled data are usually not available and the reasons behind are data unavailability (low prevalence diseases), data protection and data curation process.
This Research Topic will present and address solutions to work with small datasets and possible solutions to tackle all related issues. Transfer Learning has been considered one of the best techniques to work with small datasets. It is based on the idea that you can transfer knowledge from one task to another. However, more deep learning techniques can be used such as Bayesian Deep Learning or synthetic data generation as data augmentation. Alternatively, solutions to increase the dataset such as better annotation tools, possibly including Active Learning techniques or Federated Learning are also important in the context of small datasets.
Issues such as data heterogeneity and fairness are also of interest. Data heterogeneity (from different scanners and protocols used for imaging) is particularly important in the case of small datasets and to handle it, several approaches can harmonize the data or train the algorithm in order to improve performance. Fairness is fundamental for the future of Artificial Intelligence and awareness of biases when developing models needs to increase. Solutions to mitigate such biases are also of extreme relevance.
Working with small datasets is still difficult, but possible solutions have been presented for fields outside of Radiology and Nuclear Medicine. A perspective of possible solutions in Radiology and Nuclear Medicine, comparisons, and new out-of-box solutions are sought in this Research Topic. Analysis with respect to the number of patients as well as datasets with heterogeneous or imbalanced datasets is welcome in this Research Topic. Furthermore, analysis of biases and fairness in the context of small datasets is highly appreciated.
We are particularly interested in the articles that cover the following themes:
• Federated Learning
• Transfer Learning
• Bayesian Deep Learning
• Regularisation techniques for Deep Learning
• Synthetic Data Generation
• Unsupervised Learning
• Active Learning
• Fairness in data and models
• Data Heterogeneity
• Comparison of Radiomics and Deep Learning strategies
• Comparison of different Deep Learning strategies
Keywords:
small datasets, artificial intelligence
Important Note:
All contributions to this Research Topic must be within the scope of the section and journal to which they are submitted, as defined in their mission statements. Frontiers reserves the right to guide an out-of-scope manuscript to a more suitable section or journal at any stage of peer review.