Global Radiomic Features from Mammography for Predicting Difficult-To-Interpret Normal Cases.

Area of Science:

• Radiology
• Medical Imaging
• Artificial Intelligence in Medicine

Background:

• Mammography interpretation involves subjective difficulty, particularly for normal cases.
• Identifying difficult-to-interpret normal cases is crucial for improving reader performance and reducing diagnostic errors.

Purpose of the Study:

• To investigate the predictive capability of global radiomic features (GRFs) extracted from mammograms for identifying difficult-to-interpret normal cases (NCs).

Main Methods:

• Extracted 34 handcrafted GRFs from 239 normal mammograms using lattice- and squared-based approaches.
• Developed three random forest classifiers (CC, MLO, and CC+MLO) to distinguish between difficult- and easy-to-interpret NCs.
• Validated model performance using leave-one-out cross-validation and assessed using AUC, with statistical comparisons via DeLong test.

Main Results:

• The combined CC+MLO model showed higher performance (0.71 AUC) compared to individual CC and MLO models (0.66 AUC), though not statistically significant.
• Six GRFs were identified as valuable predictors for difficult-to-interpret NCs.
• Twenty radiomic features significantly differed between difficult- and easy-to-interpret NCs (p < 0.05).

Conclusions:

• Global radiomic features extracted from mammograms demonstrate potential in predicting difficult-to-interpret normal cases.
• Radiomic analysis offers a quantitative approach to assess mammogram interpretability, potentially aiding in quality control and reader training.