Electronics and Communication Engineering Department

SVNIT Surat

Project presentation entitled

Breast Cancer Classification using
Deep Learning
Presented by: Jinal Parmar (U18EC108),
Parth Makwana (U18EC099) & Timir Patel (U18EC038)
Guide: Dr. Jigisha N Patel
1
Outline

● Introduction
● Motivation & Objective
● Breast cancer
● Literature Survey
● Proposed Work
● Simulation/ Demo
● Results & Analysis
● Summary & conclusion
● Futurescope

2
 Introduction

● Breast cancer has ranked

number one cancer among
Indian females with age
adjusted rate as high as 25.8
per 100,000 women
● Breast cancer effective
treatement - early
detection
● AI- ML ?
Fig.1 Machine learning in healthcare
3
Motivation and Objective

● In 2019, 281,550 of women’s new cases of invasive breast

cancer are expected to be diagnosed in the U.S and 49,290
deaths were estimated.
● Detection at early stage for betterment.
● Doctors -70% accuracy (Better solution: ML)
● high-resolution histopathology pictures as input to existing
models
● Determine cancer is benign or malignant

4
 Breast Cancer

● Breast cancer diagnosis -

breast exam,
mammograms,
histopathology(biopsy)
etc
● Classification?
● Benign - non cancerous Fig.2 Benign Fig.3 Malignant
● Malignant - cancerous

5
 Literature Survey - BCC Models
Table.1 BCC Models
Author Year Technique used Dataset Findings Accuracy
S. Karthik, R. Srinivasa 2018 DNN WDBC Dataset DNN handles large 98.62%
Perumal datasets efficiently
compares to ML
algorithms

Saad Awadh Alanazi 2021 CNN IDC dataset CNN uses filters and 87%
convolution layers to
improve the model
computationally.

Teresa Araujo 2017 CNN+SVM H&E biopsy images CNN as a feature 83.3%
extractor and SVM as a
classifier
Faezehsadat Shahidi 2020 Inception- Resnet-V2 BreakHis Comparison based study 97.13%
on different transfer
learning models

Spanhol et al. 2016 Alexnet and CNN BreakHis extraction of patches 84.6%
obtained randomly or by a
sliding window mechanism
for high resolution images

6
Literature Survey - BCC Models contd.
Table.1 BCC Models

Author Year Technique used Dataset Findings Accuracy

Gour et. al. 2020 Customized ResHist BreakHis Model learns rich and 84.34%
[152- Residual Learning‐ discriminative features of
CNN] images, data
augmentation technique
(stain normalization,
image patches generation
& affine transformation)

Majid Nawaz 2018 Densenet BreakHis Modified it to deal with 95.4%

histopathology images

Sharma and Mehra 2020 VGG16+SVM BreakHis Combination of ML & 93.9%

Transfer learning with
augmentation

Han et al. 2017 CSDCNN BreakHis feature learning manner 93.2%

using prior knowledge of
class structure on
histopathological images

7
Proposed Work - Workflow

Fig.4 Workflow of project

8
Proposed Work - Dataset

● Compared to normal mammograms (expensive, have less features),

histopathological images give more details of cancerous tissues.

● BreakHis Dataset (3.9 GB)

– 9,109 microscopic images of breast Table.2 BreakHis Dataset
tumor tissue collected from 82 Magnification Benign Malignant Total
patients using diff. magnifying
40X 652 1,370 1,995
factors
(40X, 100X, 200X, and 400X). 100X 644 1,437 2,081
– To date, it contains 2,480 benign 200X 623 1,390 2,013
and 5,429 malignant samples. 400X 588 1,232 1,820
● Dataset preprocessing
Total 2,480 5,429 7,909
- rotation & flipping, 9:1
9
JinNet (Customized Model)

● Model Architecture
● consists of 7 blocks
● B1: input layer
● B2: 2conv + maxpooling
+ 1dropout (64 kernels)
● B3: similar B2 (128
kernels)
● B4: 2conv + maxpooling
(256 kernel)
● B5: flatten layer
● B6: 1dense + 1dropout
● B7: 1dense(sigmoid)

Fig.5 JinNet Architecture

10
Proposed Work – JinNet parameters

● Loss function

● Optimizer

● Early stopping for training

11
SIMULATIONS/DEMO

12
RESULTS AND ANALYSIS

13
Results & Analysis – TL Models

● InceptionResnet ● InceptionV3

Training Validation Validation Training Validation Validation

Accuracy Accuracy Loss Accuracy Accuracy Loss
50.28 48 79.25 70 0.429
0.7193 14
Results & Analysis – TL Models

● MobileNet ● Xception

Training Validation Validation Training Validation Validation

Accuracy Accuracy Loss Accuracy Accuracy Loss
87.35 86 0.37 90 86 0.414
15
 Results & Analysis – TL Models

● NasNet ● VGG16

Training Validation Validation Training Validation Validation

Accuracy Accuracy Loss Accuracy Accuracy Loss
91.68 91 0.2903 99.89 91 0.3124
16
 Results & Analysis – TL Models

● VGG19 ● DenseNet121

Training Validation Validation Training Validation Validation

Accuracy Accuracy Loss Accuracy Accuracy Loss
100 94 0.3347 97.78 95 0.289 17
Results & Analysis - JinNet

● With maxpooling ● With average pooling

Training Validation Validation Training Validation Validation

Accuracy Accuracy Loss Accuracy Accuracy Loss
91 96.06 0.18 86.9 93.33 0.371
18
Result & Anaylsis - JinNet vs TL models

Table.3 JinNet vs TL models

MODEL Training Validation Validation

Accuracy Accuracy Loss
InceptionResnet 50.28 48 0.7193
InceptionV3 79.25 70 0.429
MobileNet 87.35 86 0.37
Xception 90 86 0.414
Nasnet 91.68 91 0.2903
VGG16 99.89 91 0.3124
VGG19 100 94 0.3347
Densenet 97.78 95 0.289
JinNet 91 96.06 0.18

19
Result & Anaylsis - JinNet vs other
authors
Table.4 JinNet vs other authors

MODEL Accuracy (%)

Spanhol et al. [36] 84.6
Majid Nawaz [35] 95.4
Gour et. al. [37] 84.34
Phu T. Nguyen [42] 73.68
Sharma and Mehra 93.9
[38]
Han et al. [34] 93.2
JinNet 96.06
20
SUMMARY & CONCLUSION

● Need? Early detection

● Literature survey discuss ML, DL,TL many more
● Proposed work
● JinNet – a light weight model (12 layer
architecture)
● Results: 96.06%
● LR: 0.001
● Splitting ratio 9:1
● Optimizer: Adam
21
FUTURE SCOPE

● Different types of data set with large size

● Data pre processing - RGB to gray scale images

● Deployment of project – mobile apps purpose

22
 Thank You 
Presented by: Jinal Parmar (U18EC108),
Parth Makwana (U18EC099) & Timir Patel (U18EC038)
Guide: Dr. Jigisha N Patel

23