Detail View

One of the best ways to prevent cancer deaths is accurately detecting and examining lesions in early. Radiologists and pathologists play a vital role in making the final diagnosis. Radiologists find out a number of possible lesion candidates and assess the lesion candidates more precisely to select highly suspicious lesions. The pathologist then analyzes the selected lesions through biopsy and ultimately determines the malignancy of lesions. The main problem of this process is a considerable amount of effort and countless hours is required. Therefore, there are needs for methods for rapid and precise diagnosis of cancer. Since the success of deep learning on computer vision field, deep learning can assist radiologists and pathologists to diagnosis cancers. I mainly have conducted three deep learning applied researches for medical image analysis and they follow the diagnosis process that radiologists and pathologists do.


In the first study, we introduce a research on detecting of masses in mammograms using a state-of-the-art one-stage object detector based on a deep convolutional neural network. To validate our mass detection method performance in diverse clinical cases of mammogram data, we conduct several experiments using the public dataset INbreast and the in-house dataset GURO.


In the second study, we conduct a research on classifying lung nodules in CT scans using three-dimensional deep convolutional neural networks with a checkpoint ensemble method. We use 3D deep convolutional neural networks with shortcut connections and dense connections to capture 3D features. In addition, we use an alternative ensemble method called the checkpoint ensemble method to boost performance. 


In the last study, we introduce an automatic nuclei segmentation method based on deep convolutional neural networks for histopathology images. We use a deep convolutional Gaussian mixture color normalization model to normalize the color of histopathology images. To segment nuclei, we use Mask R-CNN which achieves state-of-the-art object segmentation performance in the field of computer vision. In addition, we perform multiple inference as a post-processing step to boost segmentation performance. 

1.   Introduction 1
2.  Detection of masses in mammograms using a one-stage object detector based on a deep CNN 5
 2.1    Backgrounds and Problem Definition 5
 2.2    Method 8
  2.2.1 Model description 8
  2.2.2 Training and Inference Procedure 11
  2.2.3 Transfer Learning 12
 2.3    Experiments and Results 13
  2.3.1 Data 13
  2.3.2 Experimental Setups 14
  2.3.3 Results 15
  2.3.4 Discussion 21
3.   Classification of lung nodules in CT scans using three-dimensional deep CNNs with a checkpoint ensemble method 25
 3.1    Backgrounds and Problem Definition 25
 3.2    Method 29
  3.2.1 Layer Connection 29
  3.2.2 Model Description 30
  3.2.3 Ensemble 33
 3.3    Experiments and Results 36
  3.3.1 Data 36
  3.3.2 Pre-processing 36
  3.3.3 Evaluation Metric 37
  3.3.4 Result 37
4.  An  automatic  nuclei  segmentation  method  based  on  deep  CNNs for histopathology images 42
 4.1    Backgrounds and Problem Definition 42
 4.2    Method 45
  4.2.1 Overview 45
  4.2.2 Pre-processing 46
  4.2.3 Color Normalization 46
  4.2.4 Nuclei Segmentation 50
  4.2.5 Post-processing 52
 4.3    Experiments and Results 53
  4.3.1 Data 53
  4.3.2 Results 54
  4.3.3 Discussion 60
5.   Conclusions 63
Bibliography 65