Sample-Size Determination Methodologies for Machine Learning in Medical Imaging Research: A Systematic Review.

Area of Science:

• Medical Imaging
• Machine Learning
• Data Science

Background:

• The optimal training sample size for machine learning (ML) models in medical imaging is frequently undetermined.
• This uncertainty impacts model reliability and generalizability.

Purpose of the Study:

• To conduct a descriptive review of existing sample-size determination methodologies for ML in medical imaging.
• To propose recommendations for future research and standardization in the field.

Main Methods:

• A systematic literature search was performed using Medline and Embase databases.
• Keywords included "machine learning," "image," and "sample size," with articles published between 1946 and 2018 included.
• Data on ML tasks, sample sizes, and train-test pipelines were extracted for qualitative analysis.

Main Results:

• Out of 167 identified articles, 22 were included for analysis.
• Only 4 studies focused on sample-size determination methodologies; 18 explored sample size effects on model performance.
• Methods varied, including pre hoc model-based and post hoc curve-fitting approaches, with significant variability in testing and reporting.

Conclusions:

• There is a significant lack of research on sample-size determination methodologies for ML in medical imaging.
• Standardization of reporting practices is crucial.
• Future work should focus on developing and streamlining both pre hoc and post hoc sample size determination methods.