Improving Cell Instance Segmentation in Scanning Electron Microscopy via Semantic Image Synthesis

Understanding tumour tissue architecture is essential for advancing new therapies and necessitates accurate cell instance segmentation in serial scanning electron microscopy (SEM) images. While artificial intelligence (AI)-based methods are effective for this task, errors persist in poor-quality cell contours, requiring manual gap filling. This study addresses cell labelling and contour segmentation by supplementing a standard segmentation technique (nnU-Net) by a specialized cell contour filling network. To train the latter, our approach utilizes a generative adversarial network (GAN)-based image synthesis method to generate realistic disturbed contour maps from randomized coarse cell contour semantic maps (CCSM). Subsequently, cell labelling was produced using the enhanced contours. Our approach was evaluated on private serial block face-SEM (SBF-SEM) images of hepatoblastoma tissues, achieving accurate cell labelling (70.8% labelling) and cell delineations (centerline Dice (clDice)= 0.91). This outperforms the state-of-the-art Cellpose algorithm (29.2% labelling, clDice= 0.34), opening the path to bioarchitecture investigations.