Towards reliable cardiac image segmentation: Assessing image-level and pixel-level segmentation quality via self-reflective references.

Cardiac image segmentation is a fundamental step in cardiovascular disease diagnosis, where many deep learning models have achieved promising performance. However, when deploying these well-trained models for real clinical usage, the network will inevitably produce inferior results due to domain shifts, motion artifacts, etc. How to avoid the potential poor-quality segmentations involved in clinical decision making is crucial for reliable computer-aided cardiac disease diagnosis. To this end, we develop a quality control method to identify failure segmentations by measuring their qualities, and report them to physicians for professional opinions. In specific, we propose a reference-based framework to assess the image-level quality (i.e. per-class Dice) for overall evaluation and pixel-level quality (i.e. pixel-wise correct map) to locate mis-segmented regions. Following previous works, we create informative references first, and investigate their relative relationships (e.g. differences) to the inputs to expose segmentation failures. However, we generate and leverage the references in different ways. We instantiate the references by recovering input images from segmentations by a self-reflective reference generator. If the segmentation is of good quality, the reference (i.e. the reconstructed image) will be close to the input image, and the inconsistency between them would be a good indicator to deduce the qualities. To effectively explore these inconsistency, we employ a difference investigator equipped with semantic class-aware compactness constraint to force the correctly-segmented features more separable to the wrongly-segmented ones. The experiments on ACDC and MSCMR datasets demonstrated our method could effectively capture segmentation failures, and the results on low-quality (Dice∈[0,0.6)), medium-quality (Dice∈[0.6,0.8)) and high-quality (Dice∈[0.8,1.0)) segmentations showed satisfying robustness of our method.