Context-aware augmentation for liver lesion segmentation: shape uniformity, expansion limit and fusion strategy

Data augmentation with context has been an effective way to increase the robustness and generalizability of deep learning models. However, to our knowledge, shape uniformity, expansion limit, and fusion strategy of context have yet to be comprehensively studied, particularly in lesion segmentation of medical images.

To examine the impact of these factors, we take liver lesion segmentation based on the well-known deep learning architecture U-Net as an example and thoroughly vary the context shape, the expansion bandwidth as well as three representative fusion methods. In particular, the context shape includes rectangular, circular and polygonal, the expansion bandwidth is scaled by a maximum value of 2 compared to the lesion size, and the context fusion weighting strategy is composed of average, Gaussian and inverse Gaussian.

Studies conducted on a newly constructed high-quality and large-volume dataset show that (I) uniform context improves lesion segmentation, (II) expanding the context with either 5 or 7 pixels yields the highest performance for liver lesion segmentation, depending on the lesion size, and (III) an unevenly distributed weighting strategy for context fusion is appreciated but in the opposite direction, depending on lesion size as well.

Our findings and newly constructed dataset are expected to be useful for liver lesion segmentation, especially for small lesions.