Evolutionary Knowledge Update for Intracranial Region Segmentation in TOF-MRA: Self-Training from Few Labeled Cases

An evolutionary knowledge update framework—self-training with confidence-gated pseudo-labels, elastic weight consolidation, and a rollback rule—was implemented for intracranial region segmentation in Time-of-Flight MR Angiography (TOF-MRA). A DeepLabV3+ model trained on 16 labeled cases consumed 1000 unlabeled examinations from the Japan Medical Image Database over ten generations; six configurations on two separately trained seeds gave eight runs, each evaluated on a four-case training holdout that also drove candidate selection. Seven runs produced no candidate above their seed; the eighth retained a model +0.0007 above it—not distinguishable from re-training stochasticity and not matched by consistent improvement in task-relevant maximum-intensity-projection measures. The rollback rule itself did not protect the deployed model: acceptance is tested against the highest recorded score, whereas what acceptance replaces is the stored checkpoint. With a tolerance of 0.005, in the permissive Naive configuration, the two separated—8 of 10 and 10 of 10 degraded candidates were admitted, and the deployed model fell by 0.0046 and 0.0033 while the recorded best never moved; with a tolerance of zero, the two moved together. The confidence score gating pseudo-labels is confounded by foreground fraction (r = −0.370) and on labeled slices tended to rank less accurate model-generated labels higher; the Dice similarity coefficient (DSC) aggregation convention alone moves the absolute score of identical predictions by 0.0152. Acceptance criteria for continual updating must be stated in terms of the checkpoint that is retained.