Table 4.

Ablation study on the contribution of each module.

Configuration	Dice (%)	IoU (%)	Params (M)	Latency (ms)
Progressive addition on BUSI dataset
Baseline (U-Net Structure)	79.2 ± 0.9	65.5 ± 1.0	31.0	18.7
+ Lightweight Encoder	83.5 ± 0.7	71.7 ± 0.8	0.75	9.8
+ Attention Module	85.8 ± 0.6	75.1 ± 0.7	0.83	11.3
+ Dynamic Inference	87.1 ± 0.6	77.2 ± 0.6	0.83	11.5
Full CEA-Net	88.3 ± 0.5	79.2 ± 0.6	0.83	11.8
Component removal on TN3K dataset
w/o Dynamic Inference	85.1 ± 0.5	74.2 ± 0.6	0.83	11.6
w/o Attention Module	84.3 ± 0.6	72.8 ± 0.5	0.75	9.8
w/o Lightweight Encoder	81.8 ± 0.7	69.2 ± 0.8	31.0	18.3
Full CEA-Net	86.9 ± 0.4	77.0 ± 0.5	0.83	11.8

All latency measurements conducted on NVIDIA RTX 3090 GPU with fixed 4-stage decoding to ensure fair comparison across configurations. Dice and IoU are reported as mean ± standard deviation over five-fold cross-validation.