AI analysis of routine CT scans predicts who benefits from advanced breast cancer therapy

Breast cancer is the leading cause of cancer death among women worldwide, and survival falls steeply once it spreads. Five-year survival is 31.9% for distant disease versus 86.7–99.6% for loco-regional disease. Hormone receptor-positive tumors often become resistant to endocrine therapy. mTOR inhibitors such as everolimus have not generally improved overall survival, possibly because they upregulate insulin-like growth factor (IGF) signaling. Xentuzumab, an antibody that neutralizes IGF ligands, is one attempt to overcome this. Tools that show early whether a regimen is working would speed drug development and spare patients ineffective treatment.

This retrospective analysis used data from a randomized phase Ib/II trial (NCT02123823, May 2014–October 2016) of exemestane and everolimus with or without xentuzumab. Of 140 enrolled patients, 106 had interpretable contrast-enhanced CT scans at baseline and week 8. The other 34 were excluded because source images were not transferred. Of the 106, 51 received the three-drug combination and 55 received the two-drug regimen. Mean age was 59.7 years (SD 9.3), and 71.7% were white. Radiologists semi-automatically segmented every detectable lesion and assessed response by RECIST 1.1.

The team tested eight imaging biomarkers previously developed in five studies of lung cancer, melanoma and colorectal cancer. Those studies covered 1,866 training and 1,906 validation patients. The biomarkers were liver tumor volume, overall tumor volume, and two radiomics features of tumor heterogeneity (Components 4 and 6), each measured at baseline and as the change by week 8. Patients were grouped as no clinical benefit (Group A: 28, all progressive disease) or clinical benefit (Group B: 78, including 66 stable disease, 11 partial responses and 1 complete response).

Seven of eight biomarkers significantly distinguished the groups. Only the week 8 change in Component 6 failed. Liver tumor volume and its change had the highest AUCs, followed by the change in overall tumor volume and baseline Components 4 and 6. Baseline liver volume reached an AUC of 79.0% (95% CI 69.3–88.7), with 85.7% specificity and 68.0% sensitivity at its optimal threshold. Group B had lower baseline and follow-up liver and overall tumor volumes, and showed marked changes in tumor heterogeneity by week 8.

An ancillary AI model used 40 variables: 3 clinical and 37 imaging features, including the number of lesions on baseline bone scans. The strongest predictors of best overall response were changes in liver tumor volume and overall tumor volume during treatment, and the number of osteoblastic lesions on baseline bone scans. This is the group's first attempt at a multimodal signature combining CT and bone scans, and it supports transferring markers across cancer types and treatments.

The findings are a proof of concept and not ready for clinical use. The analysis was retrospective and drawn from a single trial. About a quarter of patients were excluded for missing imaging, and most of the clinical-benefit group had stable disease rather than tumor shrinkage. The authors call for further validation before pan-cancer and pan-treatment application. Three authors are affiliated with Boehringer Ingelheim.