Increasing interest in the development and validation of quantitative imaging biomarkers for oncologic imaging has in recent years inspired a surge in the field of artificial intelligence and machine learning. Initial results showed promise in identifying potential markers of treatment response, malignant potential, and prognostic predictors, among others; however, while many of these early algorithms showed the optimistic ability to separate pathologic states on “in-house” datasets, it was often the case that these classifiers generalized poorly on external validation sets and thus were of limited utility in the clinical setting. This issue was additionally compounded by the frequent use of data filtering and feature selection techniques in many studies to further bolster the machine learning results in limited case scenarios, thereby biasing the overall fit and further reducing generalizability.
In an attempt to address some of these concerns and to support large-scale machine learning efforts in oncologic imaging, multiple initiatives have been pursued to make large, anonymized imaging datasets publicly available, including The Cancer Imaging Archive (TCIA) and the Multimodal Brain Tumor Segmentation Challenge (BraTS), to name a few. The hope was that these datasets would assist with appropriately-focused subject accrual, whereby the large numbers of data inputs require sufficiently large sample sizes to ensure adequate study power as well as serve as reference-standard external validators for “in-house” classifiers.
Furthermore, significant work has been done to identify robust (i.e. stable across different imaging settings), reproducible (i.e. stable across different scanners), and repeatable (i.e. stable across time) metrics, and to develop rigorous post-processing and harmonization techniques in an attempt to ensure signal stability across different scanners, institutions, and time points.
This Research Topic aims to promote advances in artificial intelligence and machine learning geared toward clinically-applicable machine learning algorithms in oncologic imaging with a specific focus on bone and soft tissue tumors. It also seeks to highlight manuscripts promoting advances in sophisticated post-processing techniques within the realm of reducing signal variability across scanners, institutions, and time points. To that end, we especially welcome articles that demonstrate the generalizability and external validity of their machine-learning approaches.
Authors are encouraged to consider external, cross-institutional validation sets, as well as the utilization of high-quality publicly available anonymized datasets. Manuscripts pursuing validation of existing algorithms or approaches may also be considered. Perspective Articles addressing common pitfalls and considerations in emerging machine learning literature may also be considered on a case-by-case scenario.
Increasing interest in the development and validation of quantitative imaging biomarkers for oncologic imaging has in recent years inspired a surge in the field of artificial intelligence and machine learning. Initial results showed promise in identifying potential markers of treatment response, malignant potential, and prognostic predictors, among others; however, while many of these early algorithms showed the optimistic ability to separate pathologic states on “in-house” datasets, it was often the case that these classifiers generalized poorly on external validation sets and thus were of limited utility in the clinical setting. This issue was additionally compounded by the frequent use of data filtering and feature selection techniques in many studies to further bolster the machine learning results in limited case scenarios, thereby biasing the overall fit and further reducing generalizability.
In an attempt to address some of these concerns and to support large-scale machine learning efforts in oncologic imaging, multiple initiatives have been pursued to make large, anonymized imaging datasets publicly available, including The Cancer Imaging Archive (TCIA) and the Multimodal Brain Tumor Segmentation Challenge (BraTS), to name a few. The hope was that these datasets would assist with appropriately-focused subject accrual, whereby the large numbers of data inputs require sufficiently large sample sizes to ensure adequate study power as well as serve as reference-standard external validators for “in-house” classifiers.
Furthermore, significant work has been done to identify robust (i.e. stable across different imaging settings), reproducible (i.e. stable across different scanners), and repeatable (i.e. stable across time) metrics, and to develop rigorous post-processing and harmonization techniques in an attempt to ensure signal stability across different scanners, institutions, and time points.
This Research Topic aims to promote advances in artificial intelligence and machine learning geared toward clinically-applicable machine learning algorithms in oncologic imaging with a specific focus on bone and soft tissue tumors. It also seeks to highlight manuscripts promoting advances in sophisticated post-processing techniques within the realm of reducing signal variability across scanners, institutions, and time points. To that end, we especially welcome articles that demonstrate the generalizability and external validity of their machine-learning approaches.
Authors are encouraged to consider external, cross-institutional validation sets, as well as the utilization of high-quality publicly available anonymized datasets. Manuscripts pursuing validation of existing algorithms or approaches may also be considered. Perspective Articles addressing common pitfalls and considerations in emerging machine learning literature may also be considered on a case-by-case scenario.