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

Mitochondria are the organelles that generate energy for cells. Many studies have suggested mitochondrial dysfunction or impairment may be related to cancer and other neurodegenerative disorders such as Alzheimer’s Parkinson’s diseases. Therefore, morphologically detailed alterations in mitochondria 3D reconstruction of highly demanded research problems performance clinical diagnosis. Nevertheless, manual segmentation over electron microscopy volumes is not a trivial task. This study proposes two-stage cascaded CNN architecture achieve automated segmentation, combining merits top-down bottom-up approaches. For approaches, conducted on objects’ localization so delineations contours can more precise. However, combinations 2D from approaches inadequate perform proper without information connectivity among frames. On hand, approach finds coherent groups pixels takes into account avoid drawbacks approach. many small areas share similar pixel properties with become false positives due insufficient localization. In proposed method, detection carried out multi-slice fusion first stage, forming cues. Subsequently, second stage learns under supervision cues stage. Experimental results show structure alleviates both which significantly accomplishes better expedites analysis.