Multiscale and Hierarchical Feature-Aggregation Network for Segmenting Medical Images

We propose an encoder–decoder architecture using wide and deep convolutional layers combined with different aggregation modules for the segmentation of medical images. Initially, we obtain a rich representation of features that span from low to high levels and from small to large scales by stacking multiple k × k kernels, where each k × k kernel operation is split into k × 1 and 1 × k convolutions. In addition, we introduce two feature-aggregation modules—multiscale feature aggregation (MFA) and hierarchical feature aggregation (HFA)—to better fuse information across end-to-end network layers. The MFA module progressively aggregates features and enriches feature representation, whereas the HFA module merges the features iteratively and hierarchically to learn richer combinations of the feature hierarchy. Furthermore, because residual connections are advantageous for assembling very deep networks, we employ an MFA-based long residual connections to avoid vanishing gradients along the aggregation paths. In addition, a guided block with multilevel convolution provides effective attention to the features that were copied from the encoder to the decoder to recover spatial information. Thus, the proposed method using feature-aggregation modules combined with a guided skip connection improves the segmentation accuracy, achieving a high similarity index for ground-truth segmentation maps. Experimental results indicate that the proposed model achieves a superior segmentation performance to that obtained by conventional methods for skin-lesion segmentation, with an average accuracy score of 0.97 on the ISIC-2018, PH2, and UFBA-UESC datasets.