Semantic Segmentation With Multiple Contradictory Annotations Using a Dynamic Score Function
Pooya Esmaeil Akhoondi, Mahdieh Soleymani Baghshah · IEEE Access · 2023
Semantic Segmentation aims to partition an image into separate regions where each region conveys certain valuable information. In recent years, various deep learning models have achieved high performance in this task. However, when several ground truth segmentations are available, aggregating the information of these segmentations into a single ground truth becomes a crucial pre-processing step. This task can become challenging when the segmentations are contradictory and the existing classes in the segmentations are imbalanced. An elegant example is the grading of Prostate Cancer in the Gleason 2019 Challenge dataset. This dataset provides six annotations from expert pathologists for each image. The high inter-observer variability among pathologists has led to conflicting annotations for biopsy images. In addition, the low diversity of biopsy tissue patterns for all Gleason grades has also resulted in an imbalanced dataset. Previously proposed methods include the Majority Voting or the Simultaneous Truth And Performance Level Estimation (STAPLE) algorithm for combining the expert annotations. In this paper, we point out that the outputs of these algorithms discard the semantics of the image in regions of high contrast which highly dissipates the performance of deep learning models trained on these ground truths. Moreover, we propose a dynamic score function to effectively solve the diversity among the annotations and balance the Gleason grading among the annotations in terms of variability and quantity. We further train a Pyramid Scene Parsing network on the final ground truth annotations and achieve a performance higher than the participating teams in the challenge.