Table 1.
Summary of tissue samples and test performance in both training and validation datasets.
Table 2.
Summary of selected MRI-based texture features to optimize CV training accuracy.
Fig 1.
ML-based MRI invasion maps show tumor-rich (>80% tumor nuclei) extent throughout ENH and BAT.
(A,B,C,E) Biopsy locations within the non-enhancing BAT zone (green dots, arrows) on T1+C (A,D) and T2W (B,E) images correspond with high-tumor (>80% tumor nuclei) and low-tumor (<80% tumor nuclei) tissue samples on histologic analysis. (C,F) Color overlay maps with manual tracings (green) around BAT show the probability (range 0–1) of tumor-rich (red) vs tumor-poor (green/blue) content, based on ML analysis and multi-parametric MRI in 60 training biopsies and 22 validation biopsies. The maps show correspondence between tumor-rich (B, red) and tumor-poor (D, blue/gray) biopsy samples.
Fig 2.
ML-based model improves tumor-rich biopsy delineation compared with CE-MRI.
(A) ML-based MRI texture model in the full dataset (n = 82, Table 1) shows higher positive predictive values (PPV) (66.7% in BAT, 81.3% in ENH) for recovering tumor-rich samples compared with CE-MRI (21.2% in BAT, 59.2% in ENH). These PPVs suggest that the ML-based model would help recover tumor-rich BAT samples with over three times greater efficiency compared with CE-MRI guidance. (B) ML-based MRI texture model in the subanalysis (n = 76, S2 Appendix) provides higher positive predictive values (PPV) (57.1% in BAT, 80.6% in ENH) for recovering tumor-rich samples (>80% tumor nuclei) compared with CE-MRI (13.8% in BAT, 59.6% in ENH). Based on these PPVs, the ML-based model would enable four times more efficient tumor-rich recovery from BAT compared with CE-MRI guidance.