(P 354) ROLE OF APPARENT DIFFUSION COEFFICIENT (ADC) MAP-BASED FIRST AND HIGH-ORDER RADIOMIC FEATURES IN DISCRIMINATING SACRAL CHORDOMAS AND CHONDROSARCOMAS WITH OVERLAPPING CONVENTIONAL IMAGING FEATURES

Objective: Although most chordomas arise from the midline and chondrosarcoma from the lateral pelvis, approximately 10% of chondrosarcomas occur near the midline, displaying similar imaging features to chondroma, causing a diagnostic challenge. This study aims to determine the diagnostic accuracy of ADC-based first and high-order radiomic features and two novel diffusion indices for differentiating sacral chordomas and chondrosarcomas.

Methods: A retrospective, multi-reader review was performed of 82 pelvic MRIs (42 chordomas and 40 chondrosarcomas) between 12/2014 and 9/2021, split into training (n=69) and validation (n=13) datasets. Lesions were segmented on a single slice on ADC maps derived. 10 ADC metrics were calculated: 8 first-order features (minimum, mean, median, and maximum ADC, and standard deviation, skewness, kurtosis, and entropy) and two novel indices: restriction index (RI, proportion of lesion with restricted diffusion) and facilitation index (FI, proportion of lesion with facilitated diffusion). Inter-reader reliability was assessed using the intraclass correlation coefficient (ICC). Regression trees were used to derive cut points and decision trees. Means and 95% confidence intervals of 107 radiomic features comparing chondrosarcoma vs. chordoma patients were sorted based on mean group differences.

Results: There was good to excellent inter-observer reliability for 8 of the 10 ADC metrics on the training dataset. There was a significant difference (p < 0.005) between chordomas and chondrosarcomas for RI, FI, median, mean and skewness using the training dataset. Optimal cut-points for diagnosis of chordoma were RI > 0.015; FI < 0.25; Mean ADC < 1.7 × 10-3 mm2/s; Skewness > 0.177. The optimal decision tree relied on FI. In a secondary analysis, the P-values of 107 radiomic features were grouped based on the feature class. Statistically significant differences (p < 0.00047) in chondrosarcoma vs. chordoma were found in 18 of 107 radiomic features, including six first-order and twelve high-order features from four texture classes, including Gray Level Co-occurrence Matrix (GLCM), Gray Level Dependence Matrix (GLDM), Gray Level Size Zone Matrix (GLSZM), and Gray Level Run Length Matrix (GLRLM).

Conclusion: The novel FI index reliably differentiates chordomas (FI < 0.25) from chondrosarcomas. Mean ADC provides a practical alternative to FI for routine clinical use. FI, ADC mean, skewness, and 12 high-order radiomic features can reliably differentiate chordomas from chondrosarcomas.