Explainable domain transfer of distant supervised cancer subtyping model via imaging-based rules extraction

Area of Science:

• Radiomics and Medical Imaging Analysis
• Computational Pathology
• Biomedical Data Science

Background:

• Radiomics, using image texture analysis for cancer assessment, faces limitations in clinical translation due to supervised model failures.
• Developing robust imaging biomarkers for prognosis requires advanced methods beyond traditional classification.
• Distant supervision, leveraging survival or recurrence data, offers a promising avenue for cancer subtyping.

Purpose of the Study:

• To assess and validate the domain-generality of a previously proposed Distant Supervised Cancer Subtyping (DSCS) model on Hodgkin Lymphoma.
• To address the instability and lack of reproducibility in radiomics across different centers.
• To propose and test a Random Forest-based Explainable Transfer Model (RF-ETM) for domain-invariant imaging biomarker extraction.

Main Methods:

• Validation of the DSCS model on two independent Hodgkin Lymphoma datasets from different hospitals.
• Comparison of model performance and interpretability across centers to identify radiomics instability.
• Development and application of the RF-ETM to test the domain-invariance of imaging biomarkers derived from cancer subtyping.

Main Results:

• The DSCS model demonstrated successful and consistent performance, but highlighted radiomics' across-center reproducibility issues.
• The RF-ETM confirmed the domain-generality of imaging biomarkers for cancer subtyping in validation and perspective settings.
• Decision rules extracted by the RF-ETM identified robust biomarkers and risk factors for clinical decision-making.

Conclusions:

• The DSCS model shows potential for reliable translation of radiomics into medical practice, requiring further evaluation on larger multi-center datasets.
• The proposed RF-ETM enhances the interpretability and domain-invariance of imaging biomarkers.
• This work supports the use of distant supervision and explainable models to overcome radiomics limitations in clinical oncology.