1 of 16

Lung Cancer Detection;�Comparative Study Between� Different Models in Deep Learning

​

Anand Suraj

​

Guided By:

Purvi Tripathi

Sunil M K

​

2 of 16

Abstract:

  • Early diagnosis
  • Tissue analysis by pathologists.
  • Automated detection.
  • Classification models:
  • Convolutional Neural Network (CNN)
  • Residual Network (ResNet)

3 of 16

Problem Statements:

  • Appropriate deep learning models.
  • Parameters to evaluate the models.
  • Dataset selection.
  • Validation.

4 of 16

Objective:

  • Locate and isolate regions of interest.
  • Feature Extraction.
  • Classification.
  • Compare based on Metrics:
  • Accuracy
  • Efficiency

5 of 16

Existing Systems:

  • Radiology.
  • X-Rays.
  • Medical Professional Reviews
  • Positron Emission Tomography/ Computed Tomography (PET/CT).

DISADVANTAGES:

  • False positive.
  • Expensive.
  • Radiation exposure.
  • Time Consuming

​

6 of 16

Proposed System:

  • Automatic detection of cancer cells
  • Extraction of Region of Interest (ROI).
  • Classification.

ADVANTAGES:

  • Scalability.
  • Distributed.
  • Convenient.
  • Low Radiation.

​

7 of 16

Modules:

  • Datasets
  • Pre-Processing
  • Feature Extraction
  • Classification (Models Implemented)
  • Result
  • Accuracy

​

8 of 16

System Architecture Diagram:

 

 

Pre-processing

Noise removal

Image level extraction

File path extraction

Model classifier

 

Final detection

healthy

cancerous

9 of 16

Data Flow Diagram:

Input CT image

Pre-processing

Train

image

Image path extraction

Tracking & prediction

Record

 

10 of 16

Model 1:��

Convolutional Neural Network (CNN):

​

11 of 16

MODEL 2:��

  • Residual Network (ResNet):

​

Epoch

Loss

0

1.505

1

1.53

2

0.851

3

0.86

4

0.77

5

0.73

12 of 16

Classifiers:���

Serial Number

Classifier Name

Training Accuracy

Test Accuracy

1

Support Vector Classifier

0.94

0.92

2

Xgboost Classifier

1

0.92

3

Random Forest Classifier

1

0.92

13 of 16

Scope of the Project:

  • Scalability.
  • Replicable.
  • Modularity.
  • Ease of Modification.
  • Versatility.

​

​

14 of 16

Conclusion:

  • Successful implementation.
  • Variety.
  • Hybrid Models.
  • Scope for Improvement.

​

​

​

​

​

15 of 16

References:

​

​

​

​

16 of 16

​

​

​

​

​

Thank You