Assessing Breast Implant Related Bias in Imaging AI: Performance of Two Chest X-Ray (CXR)-based DL Model in Patients with Breast Implants

Purpose: We hypothesized that breast implants can impair the performance of AI tool by mimicking pulmonary/pleural abnormalities. We assessed and compared the performance of a two commercial DL models for detecting chest x-rays (CXR) findings in patients with and without breast implants (BI).

Methods/Materials: Our retrospective, IRB-approved study included 905 female patients (mean age 63±18years) with BI+ and CXRs from two community and two quaternary care hospitals. Patients with BI+ were identified from the Epic Warehouse Database (EDW) SQL server. Radiology reports were obtained from the radiology reports search engine (Nuance mPower). Next, we extracted the presence or absence of the following DL assessable findings in the radiology reports: nodules, consolidation, pneumothorax, pleural effusion, cardiomegaly, mediastinal widening, atelectasis, and calcifications. All CXRs were processed with two commercial DL models (A and B). Data were analyzed with receiver operating characteristics (ROC) analysis using area under the curve (AUC) separately for each finding. De Long analysis was used to compare performance between both groups and between AI models.

Results: DL-A had significantly higher performance than DL-B in identifying all CXR finding in BI+ patients: nodule [A: 0.60 (0.538-0.663) vs B: 0.501(0.439-0.565), P=0.024], consolidation [A: 0.732 (0.638-0.825) vs B: 0.542(0.427-0.658), P=0.002], pneumothorax [A: 0.766 (0.655-0.876) vs B: 0.478(0.379-0.578), P=0.001], pleural effusion [A: 0.890 (0.847-0.932) vs B: 0.539(0.474-0.597), P=0.001], cardiomegaly [A: 0.859 (0.774-0.943) vs B: 0.491(0.391-0.591), P=0.001], atelectasis [A: 0.679 (0.629-0.730) vs B: 0.516(0.466-0.566), P=0.030] and calcification [A: 0.685 (0.629-0.741) vs B: 0.493(0.439-0.547), P=0.031] except for mediastinal widening as no difference was noted (p=0.810).

Conclusions: There are substantial variations in CXR-DL models performance in patients with breast implants. These differences stress the need for more inclusive model trainings with greater diversity to avoid bias.Therefore, radiologists must be aware of the DL model bias in the evaluation of CXRs in patients with breast implants.