Deep Convolutional Networks And Its Application To Breast Histological Image Analysis

Modified Bloom-Richardson Grading system is highly correlated with the morphological and topological features of the nuclei in breast tumor. Therefore, accurately and fast automatic detection and segmentation of nuclei are pre-requisite steps in breast cancer computer-aided prognosis system. But because of the factors such as the diversity of pathological tissue staining, the highly complexity of image scene and pathological markers, the difference of imaging methods and the diversity of imaging instrument, cause most of the nuclei present a highly complex shape and structural characteristics, such as color non-uniformity, cells overlaps and the cohesion of cells. Therefore automated detection and segmentation of nuclei is a challenged task for medical images processing. In this case, a Convolutional Neural Network initialized Local Region based Active Contour (CNNiLRAC) model is presented for nuclei detection and segmentation. The model efficiently couples two components:CNN for accurate nuclei detection and LRAC for nuclei segmentation. In the first component, the CNN which initial weights firstly learned with Sparse Autoencoder (SAE) model is integrated with sliding windows based technique for patch-wised detection of nuclei within high resolution histological images. In the second component, the initial contour of each nuclear patch is obtained with gray-scale thresholding approach and is subsequently served as the initial contour of this patch in the segmentation component. Segmentation process of LRAC model is driven by the Gaussian distribution of local intensities in both nuclei and background regions. Contrast to three different methods, such as Bule Ratio(BR), Iterative Radial Voting(IRV) and Maximally Stable Extremal Region(MSER), the performance of CNNiLRAC model can accurately detect and segment cells. Evaluated on three different breast histological data sets, the detection accuracy of CNNiLRAC was F-measure 80%,86%, and 80% and average area under Precision-Recall curves (AveP) 77%, 82%, and 74%. The segmentation accuracy was Positive Predictive Value (PPV) 85% and 89% across two data sets which have ground truth. In terms of the model of convolution network, based on Traditional Convolution Neural Network (TCNN), this paper proposed a fast and efficient Multi-level Pyramid Convolutional Neural Network (MLPCNN). This model avoided the low speed of convolution filter in processing large-scale images, the difficulty of parameter adjustment and too long time of training. This network shared the filter weights of lower to higher using weights-shared methods to ensure that the training of CNN only occur on small size image patches. This would speed up the training of CNN network. Experiments show that, in the case of the feature dimension is relatively low, the feature extraction method of MLPCNN is more effective than traditional feature extraction methods.