Stabilized Semi-Supervised Training for COVID Lesion Segmentation

We propose a novel stabilized semi-supervised training method to solve the challenging problem of covid lesion segmentation in CT scans. We first study the limitations of current models and based on our findings we introduce a lightweight SU-Net (Small U-Net) architecture. During training we feed the CT scans in sorted order of lesion occupancy and calculate a reliability score at each epoch to determine the stopping criteria. We test the proposed method on the largest publicly available COVID CT dataset called MOSMED dataset. By harnessing around 800 un-labelled COVID CT volumes comprising 25k CT slices, we improve the segmentation accuracy by around 2-4 dice percentage points depending upon the availability of labelled training data. We also compare our method with a recently published COVID lesion segmentation method called Semi-InfNet. The proposed method outperforms Semi-InfNet model and achieves state-of-the-art covid segmentation result on MOSMED dataset.