Three artificial intelligence data challenges based on CT and ultrasound

Barriers and facilitators of artificial intelligence conception and implementation for breast imaging diagnosis in clinical practice: a scoping review

“…There were mentioned ( n = 6) international scientific challenges that contribute to strengthening the benchmarking of AI approaches by providing open and reproducible evaluation (training and testing data, evaluation metrics, and tools) that strongly encourages or imposes the sharing of computer code from participants [ 47 , 57 ].…”

“…To address these challenges, some solutions can be considered, like large datasets initiatives (F1.1) (e.g., result of competitions opened to the research community [ 57 , 69 ]), more adapted algorithmic approaches (F1.2), or the promotion of data sharing and collaboration between clinical settings by supporting regulations on patient confidentiality in favor to research (such as the generalized patient consent). These actions can contribute to having larger but also more representative and diversified datasets.…”

“…For instance, pathological findings reported in radiological reports could be associated with their exact localization in the images via user-friendly reporting tools, simplifying the data annotation and expediting its processing. In the context of creating a data challenge on US breast lesions, Lassau et al [ 57 ] observed the added value of relying on professional tools to anonymize and gather annotated data from eleven clinical centers in a centralized and an automatic way, in accordance with regulations such as GDPR. Involving and training healthcare professionals with experience in medical imaging, such as radiographers in the collection and annotation of imaging data, can be another solution to address the time-consuming data preparation and processing usually performed by radiologists.…”