cDNA Microarray Image Segmentation Based on Deep Convolutional Neural Network

Image segmentation is a crucial step for microarray image processing. Due to the diversity of microarray image quality, current unsupervised and supervised methods can't obtain better results on all datasets. To avoid the invalid pixel segmentation aroused by data dimension reduction, this paper proposes an improved deep convolutional neural network (IDCNN). To ensure the size consistency between the input and output images, an end-to-end network model is built and a padding operation is introduced on input spot block images. Meanwhile, an average sampling operation is defined to replace the fully connection layer for enhancing the uniformity between the feature plane and classification result. At last, a simulation dataset and three real datasets are prepared to conduct intensive comparative experiments with KM, FLICM and SVM. Segmentation results on simulation images prove that the average accuracy of IDCNN is a bit higher than SVM, 10 percent higher than FLICM, and about 20 percent higher than KM.