Interpretable predictions from whole-body FDG-PET/CT using parameters associated with clinical outcome

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Accurate prediction of clinical outcomes is crucial for patient care but remains challenging.
• A novel deep learning methodology was explored to predict clinical outcome-associated parameters.

Purpose of the Study:

• To evaluate a deep learning-based method using tissue-wise information for predicting clinical outcome parameters.
• To demonstrate the feasibility of automated prediction of parameters like total metabolic tumor volume (TMTV) and lesion count.

Main Methods:

• Utilized the autoPET cohort (1014 FDG-PET/CT exams).
• Extracted tissue-wise projections (bone, lean tissue, adipose tissue, air) at various angles.
• Trained a deep regression and classification framework to predict TMTV, lesion count, age, sex, and cancer diagnosis.

Main Results:

• The model achieved high accuracy in predicting TMTV (R² = 0.84) and lesion count (R² = 0.90).
• Age prediction improved with multiple projection angles (R² = 0.70).
• Sex (AUC = 1.00) and cancer diagnosis (AUC = 0.95) were predicted with high accuracy.

Conclusions:

• Tissue-wise projections enable efficient and automated prediction of clinical outcome-related parameters.
• This proof-of-concept study highlights the potential of deep learning with tissue-wise projections for future clinical outcome prediction.