Fig. 5: Four examples showcasing disease diagnosis on original and neutralized X-ray images.
From: Enhancing fairness in AI-enabled medical systems with the attribute neutral framework
The attributes for each example are as follows: a female, <60 y; b female, ≥60 y; c male, <60 y; and d male, ≥60 y. The identified findings for these examples are as follows: a infiltration; b emphysema; c nodule; and d atelectasis, effusion. Each subfigure consists of an original X-ray image and its corresponding neutralized X-ray image. The neutralized attributes are sex and age, and the modification intensity α is 0.5. The attributes of the X-ray images are identified by AI judges and human judges, while the findings of X-ray images are identified by the DDM. AI judges and DDMs report output probabilities. Human judges report the voting ratio based on the evaluations of five human judges. AT Atelectasis, CA Cardiomegaly, CO Consolidation, ED Edema, EF Effusion (EF), EM Emphysema, FI Fibrosis, HE Hernia, INF Infiltration, MA Mass, NOD Nodule, PT Pleural Thickening, PN Pneumonia, PNX Pneumothorax, NF No Finding.
