Application of Machine Learning for Mean Glandular Dose Prediction Utilizing DICOM Mammography Images

Area of Science:

• Medical Physics
• Radiology
• Artificial Intelligence

Background:

• Increasing demand for open-source data fuels AI and ML applications in medicine.
• Mammography datasets are crucial for breast cancer research, but few studies predict mean glandular dose (MGD) using AI/ML.
• Accurate MGD prediction is vital for optimizing radiation dose in mammography.

Purpose of the Study:

• To investigate the feasibility of using ML and deep ML for predicting MGD from DICOM mammogram images and associated dosimetric data.
• To evaluate various regression and neural network models for MGD prediction.
• To develop and test a deep ML fusion model combining Vision Transformer (ViT) with tabular data for MGD normalized conversion factor (CF(DgN)) prediction.

Main Methods:

• Utilized a dataset of 26,988 mammography images in DICOM format.
• Evaluated eleven regression algorithms and three neural network models using five-fold cross-validation.
• Developed a deep ML fusion model (ViT + tabular data) for CF(DgN) prediction, testing multiple feature configurations.

Main Results:

• Artificial Neural Network (ANN) sequential models showed superior performance on tabular data compared to linear and tree-based models.
• The ViT deep ML fusion model with six features achieved the best predictor performance.
• Calculated mean breast thickness of 61.37 mm and mean MGD of 1.53 mGy from the dataset.

Conclusions:

• ML and deep ML models can effectively predict MGD using dosimetric tabular data and mammography DICOM images.
• The developed deep ML fusion model demonstrates strong potential for accurate MGD prediction.
• Future work should focus on larger, more diverse datasets to further enhance ML model performance in MGD prediction.