Perspective transfer model building via imaging-based rules extraction from retrospective cancer subtyping in Hodgkin Lymphoma
Image texture analysis has for decades represented a promising opportunity for cancer assessment and disease progression evaluation, evolving over time in a discipline, i.e., radiomics. However, the road for a complete translation into clinical practice is still hampered by intrinsic limitations. As purely supervised classification models fails in devising univocal imaging- based differences in tumors with different prognosis, cancer sub- typing approaches would benefit from the employment of distant supervision, for instance exploiting survival/recurrence information. In this work, we transfer our previous model for Hodgkin Lymphoma subtyping to a multi-center study case. We evaluated model performance in two independent datasets coming from two hospitals, comparing and analyzing the results. Our preliminary data confirmed the instability of radiomics due to across-center lack of reproducibility, leading to meaningful results in one center and poorer performance in the other. We then learnt stratification rules from the first dataset via Random Forest and leveraged those rules to transfer the stratification policy onto the second dataset. In this way, on the one hand, we tested the stratification ability of cancer subtyping in a validation setting and, on the other hand, enriched the noisier dataset with valuable information, in a borrowing strength fashion. The transfer of the model resulted successful. Moreover, having extracted decision rules for cancer subtyping, we were able to draw up risk factors to be considered in clinics. The work shows the potentialities of the proposed pipeline to be further evaluated in larger multi-center datasets, with the goal of translating radiomics into clinical practice.