Utilization of deep neural network in recognition of BCR/ABL gene rearrangements in fluorescence in SITU hybridization images
Wu J, Deebajah M, Lai Z, Micale M, and Yu L. Utilization of deep neural network in recognition of bcr/abl gene rearrangements in fluorescence in situ hybridization images. Modern Pathology 2020; 33(3):1493-1494.
Background: Interphase dual-color fluorescence in situ hybridization (iFISH) has been used for identification of BCR/ABL gene rearrangements in Chronic Myeloid Leukemia (CML). Artificial intelligence, particularly deep neural network (DNN), has achieved major breakthroughs in image analysis and classification. The purpose of this study is to see if DNN can be successfully trained to recognize of BCR/ABL gene rearrangements in FISH images. Database of single-cell images (101 positive, 278 negative), and original multi-cell images (33 positive, 118 negative). The classification model was built on single-cell images and use to test the end-to-end performance only on multi-cell images. The fully automatic analysis pipeline consists of single-cell detection and single-cell classification modules (Fig 1). The detection pipeline following systematic steps; starting with the RGB images are first converted into greyscale. After which we set an intensity threshold to remove the text and apply median blurring to de-noise the image. Referring to the topological structure, we detect the closing contours and generate a bounding box on the contour region. The small and overlapping regions are removed or merged. We crop and resize the original image in the bounding box regions into a dimension of 255 and rescale the RGB value into -1 to 1 as the input to the classification network. The deep neural network architecture for single-cell classification is based on VGG, which consists of 16 layers of convolution, max pooling and fully connected operations. The network outputs a binary vector indicating positive and negative. We use Adam optimizer and Cross-entropy loss to optimize the training process. During the training time, we apply flip operation as data augmentation. Results: Our end-to-end performance matrices showed a total f1-score of 98% and recall of 98%. (Table presented) Conclusions: Our study shows the deep neural network can be trained to reliably recognize BCR/ABL gene rearrangements in FISH images with pathologist-level of accuracy.