Abstract 102P
Background
Imaging is crucial in monitoring treatment effects in soft tissue sarcoma (STS), where tumor size and intensity are key factors. Standard evaluation criteria, RECIST and Choi, have been criticized for inconsistencies and limited correlation with patient outcomes. While more accurate in predicting response, Choi criteria are labor intensive for radiologists, necessitating improved evaluation methods.
Methods
We analyzed Spearman correlations between tumor diameter, intensity and volume. Cohen’s kappa was used to measure agreement between RECIST, Choi and segmentation-based criteria (using RECIST cut-offs for volume and Choi cut-offs for intensity). Wilcoxon tests compared pre-/post-radiotherapy measurements. Correlations with pathology (% necrosis, fibrosis, viable tumor) were computed when available. We trained an AI model to segment STS in MRI and benchmarked against radiologist segmentations using a modified dice coefficient, assigning a value of 1 for correctly identifying tumor absence.
Results
High diameter vs. volume (ρ = 0.94), but weak volume vs. intensity (ρ = -0.26) and diameter vs. intensity (ρ = -0.28) correlations were seen in 270 scans (142 patients), median volume 110 cm3 (IQR: 30 - 414). RECIST-Choi, Choi-segmentation and segmentation-RECIST agreements in 178 follow-ups (144 patients) were 0.38, 0.55 and 0.32. In 121 neoadjuvant radiotherapy patients significant post-radiotherapy changes were observed in tumor intensity (p<0.001), but not in volume (p = 0.40) or diameter (p = 0.16). Pathology (available in 107 patients) showed significant correlations for volume vs. necrosis (ρ = 0.44), volume vs. fibrosis (ρ = -0.44), diameter vs. necrosis (ρ = 0.40), diameter vs. fibrosis (ρ = -0.26), intensity vs. fibrosis (ρ = -0.24) and intensity vs. viable tumor (ρ = 0.26), suggesting larger tumors show increased necrosis and decreased fibrosis while higher intensity indicates more viable tumor. The AI model scored 0.86 dice in an initial 24 patient test set and 0.62 in a new clinical 47 patient set.
Conclusions
Low agreement in STS radiologic response criteria, likely due to complementary information provided by size and intensity data, underscores the need for new methods. Our preliminary segmentation results show promise but require further refinement.
Clinical trial identification
Editorial acknowledgement
Legal entity responsible for the study
The authors.
Funding
Has not received any funding.
Disclosure
All authors have declared no conflicts of interest.