Conventional deep learning-based speech enhancement models perform inference with fixed network structure. However, due to the time-varying spectral characteristics of speech and noise, the noisy signal exhibits varying subband and segmental signal-to-noise ratios. Adapting the network structure to these variations may offer computational efficiency and interpretability. This work introduces two gating mechanisms, sample-level gating (SG) and time-frequency level gating (TFG), to dynamically skip layers within the two-stage conformer blocks of the conformer-based metric GAN network. Our experiments demonstrate competitive performance while reducing multiply-accumulate operations by 36% with SG and 48% with TFG. Furthermore, we provide insights into the model behavior by analyzing the gate activations and intermediate layer outputs. Our analysis shows that initial layers prioritize low-frequency speech regions, while later layers focus on nonspeech regions and higher frequencies.