A deep learning approach to radiation dose estimation.

Area of Science:

• Medical Physics
• Radiology
• Computational Biology

Background:

• Current radiopharmaceutical dosimetry methods lack precision, neglecting individual tissue density and radiopharmaceutical concentration.
• Existing methods often rely on crude estimations, limiting reliable dose predictions in clinical practice.

Purpose of the Study:

• To develop and validate a machine learning-based hybrid method for accurate, individualized absorbed dose prediction in radiopharmaceutical therapy.
• To overcome the limitations of current crude dosimetry techniques by incorporating detailed patient-specific data.

Main Methods:

• A hybrid deep neural network (DNN) and empirical mode decomposition (EMD) method (DNN-EMD) was developed, integrating CT density maps and SPECT data.
• The algorithm was trained using Monte Carlo (MC) simulations as a reference, with EMD analyzing density map intrinsic modes.
• Validation was performed using a leave-one-out cross-validation on a cohort of 26 subjects.

Main Results:

• The hybrid DNN-EMD method demonstrated superior performance over the standard MIRD DVK dose calculation method.
• Significantly smaller mean deviations and reduced variances in dose estimations were observed with the DNN-EMD method.
• Further reduction in variance was achieved when using intrinsic modes from EMD analysis of tissue density maps.

Conclusions:

• The proposed hybrid DNN-EMD method offers a significant advancement in individualized radiation dose prediction for radiopharmaceutical applications.
• This method provides faster and more reliable dose estimations, making it suitable for daily clinical practice.
• The findings highlight the potential of machine learning and advanced signal processing in improving nuclear medicine dosimetry.