www.nu.edu.kz
Early detection of breast cancer using deep neural network classification
models on histopathological images
Supervisor: Prof. Muhammed Fatih Demirci
Student: Almukambet Abduokhapov
Outline
INTRODUCTION RELATED WORKS
METHODOLOGY Experimental Evaluation Conclusion and Future Works
Introduction
● Breast cancer (BC) is one of the most common cancers in women and it is the major cause of cancer-related mortality in women.
● Breast cancer was diagnosed in 2.3 million women, of which 685,000 women had cancer with a fatal outcome in 2020
● Imaging methods like as diagnostic mammography (X-rays), magnetic resonance imaging, ultrasound(sonography), and thermography can help detect and diagnose BC
● A biopsy is the most accurate test to determine the cancer
Introduction
Figure 1 Stages of breast cancer
Figure 2 Representation of tumor in breast
Introduction
● Motivation:
to aid in the early identification of breast cancer and decrease human error
● Problem statement:
Determine and Develop the existing most successful deep neural networks for tumor classification(benign or malignant)
● Objectives:
Improving pre-trained deep CNN models by data augmentation techniques
Testing 6 mostly-cited pre-prepared CNN models on different datasets
Related Works
Research Methodology 40x 100x 200x 400x
A
dataset for breast cancer histopathological image classification[1]
PFTAS + SVM 0.816 0.799 0.851 0.823
Breast cancer multi-classification from histopathological images with structured
deep learning model[2]
AlexNet + Aug 0.856 0.835 0.831 0.808
A comparative study of different types of convolutional neural networks for breast cancer histopathological image Classification[3]
pre-trained CNN 0.89 0.92 0.94 0.9
Our proposed study pre-train CNN + Aug +
Opt Par 0.955 0.93 0.98 0.885
Methodology-Dataset
● BreaKHis is made up of 7909 clinically representative microscopic pictures of breast tumor tissue gathered from 82 individuals and magnified by four distinct factors (40x, 100x, 200x, and 400x)
Magnification Benign Malignant Total
40x 625 1370 1995
100x 644 1437 2081
200x 623 1390 2013
400x 588 1232 1820
Total 2480 5429 7909
Table 2:Distribution images by magnification factors and tumor type
Figure 3:Sample Images of each magnification level from BreaKHis database
Methodology-Data Augmentation
Figure 4: The steps of implementation data augmentation techniques: A - original
image, B - resized image to 224 size, C - horizontally flipped, D - rotated picture,
E - cropped image, F - changed color space transformation: brightness ,contrast , hue
and saturation, G - applied Gaussian blur , H - used equalize
Augmentation type Settings
Horizontal Flip p=0,5
Rotate p=0.5, limit=15
Random Brightness Contrast p=0.5
brightness_limit=(-0.2, 0.2) contrast_limit=(-0.1, 0.1) Random Resized Crop p=0.8
scale=(0.9, 1.1), ratio=(0.05, 1.1),
Blur p=0.3
blur_limit=(1, 3)
Table 3: The settings of Data Augmentation
Methodology-CNN models
Model Architecture
MobileNetV2 convolution layer with 32 filters, followed by 19 residual bottleneck layers
ResNet50 48 convolutional layers
1 max-pooling layers 1 average-pooling layers
Inception-v3 48 layers deep
6 convolutional layers 1 pooling layer
Inception-ResNet-v2 164-layer network
Figure 5 Inception-v3 architecture
Methodology-CNN models
● Hyperparameters:
○ Optimizer = SGD
○ Number of epochs = 12
○ Batch size = 64
○ Loss function = Binary Cross-Entropy
Figure 6 EfficientNetV2 b0 architecture[4]
Methodology - Metrics
Trained 5 models on different datasets evaluation metrics includes:
where TP - True Positive, TN - True Negative
FP - False Positive, FN - False Negative
Experimental Evaluation
● For testing set we get 20% of data set
● For training set we get 80% of data set
Experimental Evaluation
f) EfficientNetV2 b0
Experimental Evaluation
The highest results were shown by the models:
Efficientnetv2,Mobilenet-v2,Resnetv2-50 and VGG16.However, the best result was shown by Efficientnetv2, with a confident 94.5% accuracy.Overall,Efficientnetv2 has worked very well and represented top results in comparison other CNN architectures
Experimental Evaluation
Figure 7 Confusion Matrix of Efficientnetv2-b0
CNN 5 layer Pre-trained
CNN Pre-trained
CNN+aug Pre-trained CNN+aug +fine-tuning
80% 83%-84% 88%-89% 92%-98%
Conclusion
Evaluated the 5 CNN models’ performance
Compared and Trained models on different magnification factor images
A lot of experiments have been done on different data augmentation techniques and parameters of each model
Modified and Improved deep CNN architectures
Thank you for your attention!
References
[1] Fabio A Spanhol, Luiz S Oliveira, Caroline Petitjean, and Laurent Heutte. A dataset for breast cancer histopathological image classification. Ieee transactions on biomedical engineering,
63(7):1455–1462, 2015.
[2] Zhongyi Han, BenzhengWei, Yuanjie Zheng, Yilong Yin, Kejian Li, and Shuo Li. Breast cancer multi-classification from histopathological images with structured deep learning model. Scientific reports, 7(1):1–10, 2017.
[3] Farjana Parvin and Md Al Mehedi Hasan. A comparative study of different types of convolutional neural networks for breast cancer histopathological image classification. In 2020 IEEE Region 10 Symposium (TENSYMP), pages 945–948. IEEE, 2020
[4] Mingxing Tan and Quoc Le. Efficientnetv2: Smaller models and faster training. In International Conference on Machine Learning, pages 10096–10106. PMLR, 2021.