This document presents research on using convolutional neural networks (CNNs) for computer-aided diagnosis of breast cancer. The researchers trained multiple CNN models - including VGGNet, ResNet and DenseNet - on ultrasound images of breast tumors segmented from the original images. They evaluated the models' performance on three datasets and found that ensembling the predictions from several CNNs using weighted averaging achieved better diagnostic accuracy than using a single CNN model alone. The researchers conclude that tumor shape provides important diagnostic information and that ensemble learning is effective for computer-aided breast cancer diagnosis from ultrasound images.
Computer‐Aided Diagnosis of Breast Cancer Using Ensemble Convolutional Neural Networks [Ifmia2019 (no.58)]
1. Computer‐Aided Diagnosis of Breast Cancer
Using Ensemble Convolutional Neural
Networks
Department of Computer Science and Information Engineering,
National Taiwan University, Taipei, Taiwan
Speaker : Yan-wei Lee
MOST Joint Research Center for AI Technology and All Vista Healthcare
3. Introduction
• Breast Cancer
• Breast cancer is the most common malignancy in
women worldwide.
• Statistics have >40,000 deaths each year.
• If it can be detected and diagnosed early it is also
among the most curable cancer types.
3
4. Introduction
4
• Computer-Aided Diagnosis (CAD)
• Even well-trained experts may have a high variation rate.
• CAD is used to assist medical image interpretation.
• A major task for CAD systems is to classify whether a
tumor is malignant or benign.
5. Material
• Data Acquisition
• All data were collected from June 2000 to January 2018.
• Data were collected across 8 different ultrasonic
platforms, challenging the robustness of CNN.
5
6. Material
• Data Characteristics
• The data consists of a total of 1425 patients with 1687
tumors that have biopsy-proven diagnosis.
6
7. Method
• Proposed Method
• Tumor Segmentation
• The mask images generated from
the U-Net FCN model.
• CNN Architectures
• CNN Features Extraction
• Diagnosis Result
• Ensemble Method
7
Ultrasound
Image
Tumor
Segmentation
Classification
Feature
Extraction
CNN
Diagnosis Result
8. Region of Interest (ROI) Extraction
8
• The ROI of B-mode US image is manually
cropped by an expert-defined
• The bounding box was manually chosen surrounding
the tumor.
ROI
9. Tumor Image
9
• Purpose
• Shape of the breast tumor has a high correlation with
the tumor diagnostic result.
• Tumor Region Extracted
• U-Net method.
→ Conv, ReLU
→ Copy and crop
↓ Max-pooling 2x2
↑ Up-Pooling 2x2
→ Conv 1x1
Input
image
Output image
(segmented
map)
32 32
64 64
128 128
256 256
512
256 256
128 128
64 64
32 32 1
ROI Image Tumor ImageMask Image
10. CNN Architecture
10
• Purpose
• Different CNN architectures affect the performance.
• We could know which kind of CNN architecture
will be more suitable for what types of medical
image in breast cancer.
• CNNs
• VGG Net
• ResNet
• DenseNet
11. VGG Net
11
• Introduction
• VGG Net has been widely used for various
application on computer vision domain.
• Object classification
• Image segmentation
• VGG Net extracts more information by using more
hidden layers.
• 16 Layers
• 19 Layers
12. VGG-Like and VGG-16
12
• VGG-16
• We reduced the output from 1000
classes outputs to 2 classes outputs.
• 0 for benign and 1 for malignant.
• VGG-Like
• In order to improve time consuming
and memory required.
16. Ensemble Method
16
• Propose
• Get a better and more comprehensive generalized model.
• Even if a weak classifier got a wrong prediction, but whole
ensemble classifiers could correct the error back.
17. Ensemble Method
17
• Combine strategy
• Unweighted average
• Take all base machines output probability to average and output.
• Weighted average
• Each predicted probability multiplied by the weight, then average
and output.
• Weighted voting
• Count the vote and multiply the corresponding weight.
• Stacking (stacked generalization)
• Train a model for combining other models.
18. Ensemble Method
18
Base machine 1
Base machine 2
Base machine 3
Testing data
Ensemble
combine strategy
Prediction
“B/M”
Ensemble classifier
𝑝1
𝑝2
𝑝3
𝑖=1
4
𝑤𝑖 𝑝𝑖
• Three base machine with Three different data sets.
19. Result
19
• Different image data sets
• Dataset 1
• original tumor ROI.
• Dataset 2
• segmented tumor.
• Dataset 3
• tumor mask.
(Dataset 1)
(Dataset 2)
(Dataset 3)
24. Ensemble method results
24
• Combine Strategy
• Unweighted average (UA), Weighted average (WA),
Weighted voting (V), Stacking (S)
25. Conclusions
25
• Tumor shape can provide very helpful information
on diagnosis.
• The diagnostic performance of the ensemble method
was better than other models which using single
CNN architecture.