Research On Cancer Image Recognition Based On Convolutional Neural Networks

The rapid developments of medical imaging technologies and medical image recog-nition algorithms reflect the strong demand for medical information. Nowadays, med-ical images which carry a wealth of information about lesions are playing important roles in diagnosing. However, artificial recognition results of medical images are eas-ily affected by cognitive ability, subjective experience and fatigue degree. In contrast,computer recognition algorithms could get over these shortcoming, effectively improve the accuracy and stability, and reduce misdiagnosis and missed diagnosis. So research on medical image recognition algorithms is of great significance for clinical diagnosis,pathological analysis and treatment options. Convolutional neural networks are kinds of learning-based feature learning methods. Algorithms based on convolutional networks have rapidly became the first choice for medical images recognition, e.g. classification,detection and segmentation. In this thesis, image classification and semantic segmen-tation algorithms for cancer recognition are studied. The main work includes:1) CNN-SVM for microvascular morphological type recognition. Recognition of microvascular morphological type, which is closely related to the identification of esophageal cancer, is the prerequisite for the cancer diagnosis and treatment. The data-driven CNN models are more applicable for complex and variable microvascular vision patterns than manual design feature extractors. In the case of relatively small sample size, a recognition system with CNN-SVM as the core model is designed. The robust-ness of predictions on scaled and rotated images is improved with data augmentation.For classifier promotion, the generalization ability of the proposed system is enhanced by SVM in place of softmax. Compared with the extensively used handcrafted features,CNN demonstrates superior feature representation ability.2) Multi-constrained FCN for microvascular morphological type semantic segmen-tation. In this thesis, a semantic segmentation algorithm for microvascular morpho-logical type recognition is proposed. In the case of incomplete-labeling, FCN model with double-label for semantic segmentation is constructed. Combined with artificial knowledge, the roi-label indicating region of interest is extracted from the annotation information as extra constraint to guild feature learning. In framework of multi-task learning, FCN model with two target co-optimizes the shared encoder to improve the performance of semantic segmentation.3) Jointly learned FCN for semantic segmentation on cell image. Microscopic cancer cell recognition is the main content of pathological examination, and the key to cancer diagnosis. However, cancer region labeling is difficult. In this thesis, a seman-tic segmentation algorithm for identifying metastatic breast cancer regions is proposed.According to multi-task learning theory, a related classification task is designed, and the joint learning method of CNN and FCN is put forward. Making a thorough inves-tigation of the classification task, the model optimization is completed and the value of the additional dataset is verified. Within the framework of multi-task learning, the performance of segmentation task is improved indirectly by increasing the performance of classification task.In summary, the feature representation ability of convolution neural networks is improved from both aspects of model and data. Research on model involves design-ing the network structure, introducing SVM, using BN normalization and exploring the optimization algorithm based on gradient descent. Research on data involves investi-gating data augmentation, designing roi-label, and making use of additional datasets.The experimental results show that recognition tasks benefit from these improvements.