Automatic sequence identification in multicentric prostate multiparametric MRI datasets for clinical machine-learning

Area of Science:

• Radiology and Medical Imaging
• Machine Learning in Healthcare
• Prostate Cancer Diagnostics

Background:

• Organizing large multi-centric multiparametric MRI (mpMRI) datasets for prostate cancer (PCa) machine learning (ML) is time-consuming.
• Accurate sequence-type identification is essential for curating these datasets to train robust clinical ML models.

Purpose of the Study:

• To develop and validate an accurate ML method for automatic sequence-type identification in multi-centric PCa mpMRI datasets.
• To create knowledge-based heuristics to further enhance automated series classification.

Main Methods:

• Retrospective classification of prostate mpMRI studies into five series types (T2W, DWI, ADC, DCE, others).
• Training of XGBoost and CatBoost models using metadata, with 5-fold cross-validation and learning curve analysis.
• Validation using hold-out and temporal test sets, and Leave-One-Group-Out cross-validation to assess center-specific effects.

Main Results:

• High test F1-scores (>0.95 for CatBoost, >0.97 for XGBoost) achieved.
• Models demonstrated learning saturation and temporal generalization capabilities for T2W/DWI/ADC triplets.
• Performance decreased when center-specific data was excluded, particularly for CatBoost, highlighting the need for such data.

Conclusions:

• Automatic sequence-type identification using ML is feasible and enables automated data curation for PCa mpMRI.
• While models generalize temporally, optimal performance necessitates the inclusion of dataset-specific data.
• Developed heuristics can assist researchers in series classification for PCa mpMRI datasets.