Fig. 4: Scenarios 2 and 3: Prediction alignment over time under continuous acquisition shift.
From: Automatic correction of performance drift under acquisition shift in medical image classification
Each simulation is repeated 250 times, solid lines depict the average difference between sensitivity and specificity across all bootstrap samples and shaded regions denote the 5%-95% percentile bootstrap confidence interval. Plots in the left column depict the number of scans processed by scanner A and scanner B over time for each scenario. Plots in the middle column compare the evolution of the sensitivity-specificity balance over time with and without applying UPA across all scans. The right-most two columns compare the evolution of the SEN/SPC balance scanner-wise. The goal is to avoid a drift between sensitivity and specificity in the presence of a gradual acquisition shift. The proposed method successfully maintains a null difference between sensitivity and specificity over time, whereas the non-adapted model can lead to dramatic shifts in the sensitivity-specificity balance. Source data are provided as a Source Data file.
