Semi-supervised Medical Image Segmentation through Dual-task Consistency

Deep learning-based semi-supervised learning (SSL) algorithms have led to promising results in medical images segmentation and can alleviate doctors' expensive annotations by leveraging unlabeled data. However, most of the existing SSL literature tend regularize model training perturbing networks and/or Observing that multi/dual-task attends various levels information which inherent prediction perturbation, we ask question this work: explicitly build task-level regularization rather than implicitly constructing networks- data-level perturbation then for SSL? To answer question, propose a novel dual-task-consistency framework first time. Concretely, use dual-task deep network jointly predicts pixel-wise map geometry-aware level set representation target. The is converted an approximated through differentiable task transform layer. Simultaneously, introduce consistency between set-derived maps directly predicted both labeled Extensive experiments on two public datasets show our method largely improve performance incorporating Meanwhile, outperforms state-of-the-art methods.