Purpose or Case Report: Pediatric radiologists are self-taught and independently interpret artificial intelligence manuscripts. However, they might not identify the pitfalls and biases related to complex data from pediatric models.
Pediatric imaging differs from adult imaging in several instances such as sample size, growth-related disorders, disease presentation, quality standards, image distortions and motioning, radiation safety adjustment, and other challenges that can impact the development of algorithms and models. The specific metrics to measure the performance of the model must be in accordance with the data and clinical task, as it can vary within the prevalence of the disease and with the balance of the dataset, and therefore can mislead the model performance results.
Challenges related to the model development can arise from data management, data transformation, small datasets, imbalanced datasets, optimization, and type of validation.
All those pieces of information are as important as the medical aspect, however, in artificial intelligence, manuscripts will explore computational engineering and data science more than clinical knowledge. The understanding of the learning architecture, classifiers, parameter, and algorithms will allow the reader to translate engineering details into clinical meaning.
In this work, we will present pitfalls and biases during data curation, hyperparameters, weights, validation, and classification metrics; and discuss the best practices to mitigate those challenges. This work integrates our group experience with a thorough literature review of prior artificial intelligence manuscripts in pediatric radiology for the analysis and application of artificial intelligence to clinical practice.
Our goal is to improve critical thinking when interpreting scientific research in artificial intelligence for pediatric radiologists.
Methods & Materials:
Results:
Conclusions: