CEIMVEN: An Approach of Cutting Edge Implementation of Modified Versions of EfficientNet (V1-V2) Architecture for Breast Cancer Detection and Classification from Ultrasound Images

Paper ID- 85

​

Presenter:

Sheekar Banerjee

Senior AI & ML Engineer, KaleidoSoft, Zagreb, Croatia

Research Advisor, Department of Computer Science and Engineering, IUBAT

​

Co-author:

Md. Kamrul Hasan Monir

AI & ML Engineer, KaleidoSoft, Zagreb, Croatia

Contents

• Introduction
• Related Works
• Dataset
• Proposed Method
• Results & Accuracies
• Conclusion & Future Endeavors

Breast cancer is cancer that develops from breast tissue.

Image Source: Nevit Dilmen

Introduction

Breast Ultrasound Imagery

Image Source: DenseBreast info INC. and Dr. Ellen Mendelson

• Automatic breast tumor screening of mammographic images with optimal convolutional neural network (2022)

​

• Used BreakHis Dataset
• Multi-layered CNN
• Maximum 95.81 % of Accuracy

Literature Review�

• MultiNet: A deep neural network approach for detecting breast cancer through multi-scale feature fusion (2022)

​

• VGG16, DenseNet-201 and NasNet-Mobile
• Maximum 96.01 % of Accuracy

​

​

​

Literature Review (Con’t)�

• Multiclass classification of breast cancer histopathology images using multilevel features of deep convolutional neural network

​

• Used “Amassed” Dataset
• Xception’s six intermediate layers
• Maximum 96.5 % of Accuracy

Literature Review (Con’t)�

Motivation

• Implementing a novel model by modifying the cutting-edge Efficient-Net-V1-V2 architecture.

​

• Getting the most accurate possible prediction for breast cancer classification

​

• A speedy and quick prediction

​

• Testing the prediction of malignant cell with ROI Biomarker.

​

Dataset

• Live breast ultrasound images.

​

• 600 female patients, aged between 25 and 75.

​

• 780 raw images, Augmented into 4500 more replicated images.

​

• Al-Dhabyani W, Gomaa M, Khaled H, Fahmy A.: Dataset of breast ultrasound images. Data in Brief. 2020 Feb; 28:104863. DOI: 10.1016/j.dib.2019.104863.

Methodology

​

• Data Labelling

​

• Architecture for modified Efficient-Net-V1

​

• Architecture for modified Efficient-Net-V2

​

• Model training and hyper-parameter tuning

​

• Keras (Backend), Tensorflow JS and TF-Lite

Data Labelling

Masked Image

Raw Image

The architecture of our proposed modified EfficientNet-V1 (b0-b7)

The architecture of our proposed modified EfficientNet-V2 (b0-b3)

EfficientNet Model scaling for our proposed method

Modified Training:

- 5280 Images from the dataset (780 Raw, 4500 Augmented)

​

- Labeling and Annotating Images

​

• Primary Training (20 Layers)

​

• 10 Layers for Efficient-Net-V1 (B0-B7)

​

• 10 Layers for Efficient-Net-V2 (B0-B3)

​

- Re-scaling (0.1/255)

​

- Hyper-parameter Tuning

Result Analysis

Testing the prediction of the classes (benign, malignant and normal)

Result Analysis (Con’t)

Breast cancer cell detection from ultrasound image as “Benign” with 95% confidence, “Malignant” with 96% confidence and “Normal” with 99% confidence.

Result Analysis (Con’t)

Table 1. Modified EfficientNet-V1 models’ average AUC (training, validation) for proposed breast cancer classification:

Result Analysis (Con’t)

Table 2. Modified EfficientNet-V1 models’ average Loss (training, validation) for proposed breast cancer classification:

Result Analysis (Con’t)

Table 3. Modified EfficientNet-V2 models’ average AUC (training, validation) for proposed breast cancer classification:

Result Analysis (Con’t)

Table 4. Modified EfficientNet-V2 models’ average Loss (training, validation) for proposed breast cancer classification:

Result Analysis (Con’t)

Table 5. Accuracies of Prediction and Detection, Precision and Recall rate for each classes:

Limitations

• Slower Run Time
• Less amount of GPU
• Combination of more modified Neural Architecture is necessary
• Limited Memory Space of Hardware Machinery.

Conclusion & Future Endeavors

• Expand the boundaries of computer vision algorithmic approach into the field of medical image and deep learning.

​

• Focusing to take our research at further level of implementation of new cutting-edge Neural Network Architectures such as: Inception-ResNet-V2, Fast-AI, Dense-Net (121, 169, 201) etc.

Any

Question?

Thank

You!