Deep learning on classifying small lesions from small datasets for computer-aided diagnosis

Abstract

In a recent decade, the deep neural network has received high attention as a core technique for various applications in computer-aided diagnosis, such as medical image segmentation and lesion classification. As in general vision tasks, the deep neural network training or deep learning has been consistently reported to achieve the state-of-the-art performances in numerous large-scale medical image tasks, e.g. chest X-ray evaluation, lung CT nodule detection, and breast mass classification in mammography. However, it has been recently discussed that there are some unique challenges of applying deep learning in general medical image tasks, especially in the classification of lesions, masses, or tumors. First, except for the aforementioned tasks, the most lesion classification tasks suffer from small dataset problem in which the size of the provided dataset is usually insufficient to train the deep neural network. Second, despite the fact that the most of major imaging features for classifying the lesions are considered to be the texture information according to the clinical studies, the deep neural network is known to be susceptible to size bias problem in which the classifier is trained to classify the lesions according to the size and shape rather than the texture information. In this thesis, two approaches are presented to improve the learning efficiency of the deep neural networks from the perspectives of above two challenges for two lesion classification tasks

The small renal masses (SRM) and the focal liver lesions (FLL). The classification of SRM in CT images as benign angiomyolipoma without visible fat (AMLwvf) and clear cell renal cell carcinoma (ccRCC) is an important task for diagnosis and surgical planning of renal cancer patients. However, in addition to the imaging similarity between AMLwvf and ccRCC, the task has two major challenges including (1) a small dataset problem with less than 100 patient data in total, and (2) a size bias problem due to the practical (not clinically meaningful) difference between average lesion sizes of AMLwvf and ccRCC. In the thesis, we propose a deep feature classification method to distinguish AMLwvf from ccRCC in abdominal CT images. To overcome the challenges, we (1) construct the random forest classifier with the concatenated features of pre-trained deep- and handcrafted features, to increase the learning efficiency of extremely small dataset, and (2) propose the texture image patch for deep neural network input patch, to reflect the texture information for deep features and to reduce the size bias. In experiments, we confirmed that the proposed approach improved the learning efficiency compared to the conventional methods. The classification of FLLs in CT images as a benign cyst, hemangioma, and malignant metastasis is an important problem for the diagnosis of cancer stages and the treatment planning of cancer patients. However, the problem has also a number of challenges including (1) a small dataset problem with less than 1,000 training dataset, and (2) an overlap between different diseases classes due to their imaging similarity and size bias problem. In the thesis, we propose a deep learning classification method to classify cysts, hemangiomas, and metastatic FLLs in abdominal CT images. To overcome the difficulties, we (1) augment the training data using the generative adversarial network (GAN) to synthesize the novel patterns of existing real data, and (2) propose the lesion information augmented (LINA) patch to reflect both the lesion texture and boundary information. In experiments, we confirmed that the proposed GAN data augmentation and LINA patch training improved the deep learning efficiency for the given task.