Two-stage Cascaded CNN Model for 3D Mitochondria EM Segmentation

Mitochondria are the organelles that generate energy for cells. Many studies have suggested that mitochondrial dysfunction or impairment is highly related to cancer, Alzheimer's and Parkinson's diseases. Therefore, morphologically detailed alterations in mitochondria and the corresponding 3D reconstruction are highly demanded for both research analysis and clinical diagnosis. However, manual segmentation of mitochondria over 3D electron microscopy volumes is not a trivial task. In this study, a two-stage cascaded CNN architecture is proposed to achieve automated 3D mitochondria segmentation, which combines the merits of the top-down approach and the bottom-up approach in segmentation. In the first stage, the detection of mitochondria is carried out with the scheme of detection stacking, becoming the segmentation cues for localization information. Subsequently, the second stage is to perform 3D CNN segmentation that learns the voxel characteristics and 3D connectivity properties under the supervision of the detection cues from the first stage. The performance of the final segmentation results by our Model-S3 with TTA 3 reaches 0.995 in accuracy, 0.966 in dice coefficient, 0.935 in foreground IoU, and 0.965 in mean IoU. The experimental results show that the proposed framework can alleviate the problems in both top-down and bottom-up approaches and achieve the state-of-the-art performance in segmentation. The framework for automated 3D mitochondria EM segmentation is expected to facilitate the clinical analysis significantly.