CQC-Net: A Lightweight Model Utilizing Consistent Orientational Quaternion Wavelet Convolution for Skin Lesion Segmentation

Accurate and efficient segmentation is crucial for melanoma diagnosis. Recent approaches have shifted from focusing solely on spatial information to incorporating frequency information, such as via wavelet transforms, to balance performance and model complexity. While these methods have demonstrated success, they often overlook intrinsic directional properties of wavelets (vertical, horizontal, and diagonal components) and the interplay between low- and high-frequency components. To address these gaps, we propose an orientation-consistent quaternion convolution (OCQC) module and a quaternion enhanced feedforward network (QEFN), both operating in the quaternion wavelet domain. The OCQC module leverages directional properties of quaternion wavelet transforms, applying direction-specific quaternion convolutional kernels to different frequency components to avoid redundant feature learning. The QEFN uses quaternion depthwise separable convolutions (QDSC) and inverse QDSC (IQDSC) to project features into higher dimensions and reconstruct them, facilitating interaction among frequency bands. By integrating the quaternion wavelet transform (QWT), inverse QWT (IQWT), OCQC, and QEFN, we propose the consistent quaternion convolution neural network (CQC-Net). Extensive experiments show that our method achieves competitive performance while maintaining efficiency, with only 0.86 M parameters and 2.96 G Flops.