Weakly Supervised Retinal Detachment Segmentation Using Deep Feature Propagation Learning in SD-OCT Images

Most automated segmentation approaches for quantitative assessment of sub-retinal fluid regions rely heavily on retinal anatomy knowledge (e.g. layer segmentation) and pixel-level annotation, which requires excessive manual intervention and huge learning costs. In this paper, we propose a weakly supervised learning method for the quantitative analysis of lesion regions in spectral domain optical coherence tomography (SD-OCT) images. Specifically, we first obtain more accurate positioning through improved class activation mapping; second, in the feature propagation learning network, the multi-scale features learned by the slice-level classification are employed to expand its activation area and generate soft labels; finally, we use generated soft labels to train a fully supervised network for more robust results. The proposed method is evaluated on subjects from a dataset with 23 volumes for cross-validation experiments. The experimental results demonstrate that the proposed method can achieve encouraging segmentation accuracy comparable to strong supervision methods only utilizing image-level labels.