A CNN Based HEp-2 Specimen Image Segmentation and Identification of Mitotic Spindle Type Specimens

In the proposed work, an effective framework for identification of mitotic type staining patterns is demonstrated, integrated with a segmentation approach, in order to detect the autoimmune disorders using HEp-2 based cell substrates. It is shown that the segmentation approach obviates the requirement of DAPI channels in Indirect Immuno-Fluorescence (IIF) imaging process. Moreover, the segmentation is required for cell-based processing of staining patterns, which is in turn required, due to the rare appearance and occurrence of mitotic type cell patterns. The segmentation involves a pixel labeling strategy, using U-Net, a Deep Convolutional Neural Network (D-CNN) based approach. The effectiveness of such segmentation approach is shown by the subsequent performance for detection of Mitotic Spindle (MS) type staining patterns, framed as mitotic v/s non-mitotic/interphase classification problem. This classification task is effectively addressed using features extracted from a traditional filter bank and CNN based feature representation. After identification of individual MS cells in specimens, a threshold-based MS detection criteria has been used specifically for declaration of specimens, consisting of MS patterns. The current study demonstrates a comparative analysis of proposed segmentation masks, with given DAPI-based masks with encouraging results for segmentation, as well as classification, over a publicly available dataset.