This document proposes using machine learning and deep learning techniques to develop an application for predicting lung cancer from CT images. It discusses how cancer is a leading cause of death and how computer-aided diagnosis using medical images can help experts. A dataset of 1500 patient CT images is preprocessed and features are extracted using HOG. Various machine learning models including KNN, CNN, and others are trained on the images and tested to achieve up to 85% accuracy. Finally, an application is developed to allow users to select an algorithm and test images for cancer detection.

2. LUNG CANCER
DETECTION USING
MACHINE LEARNING AND
DEEP LEARNING
TECHNIQUES
BY HAMZA SHAUKAT

3. INTRODUCTION
• Cancer is one of the most deadly disease that humans have ever encountered
• Most common cause of cancer death in men and women
• Tobacco use accounts for 87% of Lung Cancer
• In 70% of lung cancer patients the disease is already spread to other organs
• Not easy to detect in early stages
• Menace for humankind

4. COMPUTER AIDED DIAGNOSIS
• With the advancement in technology computers are now used in the medical
field to diagnose different diseases by reading the information in medical images
such as CT images and MR images etc.
• Computers perform this time consuming and redundant tasks more efficiently
• These techniques have revolutionized the medical world and taken burden off
the shoulders of medical experts specially in decision making

5. OBJECTIVES
 Develop an application for the prediction of lung cancer with the help of CT images using machine
learning and deep learning techniques.
 Increase the accuracy of performance of Lung Cancer Prediction.
 Also reduce the time and cost required for various excessive Medical tests.
 Comparison of the different algorithms and analysis of our results.

6. DATASET
• THE DATASET WAS TAKEN FROM THE CANCER IMAGING
ARCHIVE
• IT CONTAINED CT IMAGES OF 1500 PATIENTS AND A CSV FILE
WHICH HAD INFORMATION ABOUT THE PATIENTS

7. PRE-PROCESSING
• THE IMAGES WERE CONVERTED FROM DICOM TO JPG FORMAT
• CORRUPTED IMAGES WERE REMOVED FROM THE DATASET
• THE IMAGES WERE RESIZED FROM (512X512) TO (256X256)
• THE IMAGES WERE CONVERTED INTO GRAYSCALE
• DIMENSION OF IMAGES WAS EXPANDED USING NP.EXPAND_DIMS()
• ALL IMAGES WERE APPENDED AND SAVED IN THE FORM OF A NUMPY
ARRAY

8. PRE-PROCESSING
• THE VARIABLE X CONTAINED THE IMAGES
• THE LIST NAMED Y IS ALS0 INITIALIZED WHICH HOLDS THE
CLASS F ALL THE IMAGES ie 0 OR 1

9. FEATURE EXTRACTION USING HOG
• HOG FEATURE DESCRIPTOR WAS USED TO EXTRACT USEFUL
FEATURES OF THE IMAGES
• HOG GETS THE MAGNITUDE AS OF THE GRADIENT AND
DIRECTION OF THE EDGES

10. MACHINE LEARNING
• ALL THE NECESSARY LIBRARIES ARE IMPORTED eg SKLEARN etc
• IMAGES ARE SPLIT INTO 80/20. 80 % FOR TRAINING THE
MACHINE LEARNING MODELS AND 20 % FOR TESTING THEM
• THIS IS DONE USING TEST TRAIN SPLIT FUNCTION

11. TRAINING
• ALL THE MACHINE LEARNING MODELS WERE TRAINED ON 80 %
OF THE DATASET
lda=LinearDiscriminantAnalysis()
lda.fit(X_train,y_train)
SIMILARLY 12 OTHER MODELS WERE TRAINED ON THIS DATA

12. SAVING TRAINED MODELS
• EACH TRAINED MODEL WAS SAVED USING PICKLE LIBRARY IN
THE FORM OF .PKL FORMAT SO IT CAN BE LATER USED FOR
LUNG CANCER DETECTION
with open('lda.pkl', 'wb') as file:
model=pickle.dump(lda, file)

13. TESTING
• THE TRAINED MODELS WERE LOADED USING PICKLE LIBRARY
• MODELS WERE TESTED ON UNSEEN DATA
• CONFUSION MATRIX AND CLASSIFICATION REPORT WAS
GENERATED FOR EVERY MODEL AND SAVED IN THE FORM OF
IMAGE AND TEXT FILE RESPECTIVELY

14. RESULTS
THE RESULTS OF ALL THE ALGORITHMS THAT I APPLIED ARE
GIVEN BELOW FOR COMPARISON AND ANALYSIS
CONFUSION MATRIX AND CLASSIFICATION REPORT ARE
PROVIDED IN THE DOCUMENTATION

15. RESULTS
COMPARATIVE ANALYSIS
S.NO CLASSIFIER PRECISION RECALL F1-SCORE ACCURACY
1 SUPPORT VECTOR 0.67 0.64 0.65 66%
2 DECISION TREE 0.66 0.66 0.66 66%
3 RANDOM FOREST 0.76 0.76 0.76 76%
4 BERNOULLI-NB 0.57 0.66 0.56 57%
5 GAUSSIAN-NB 0.59 0.58 0.58 58%
6 K-NEAREST NEIGHBORS 0.80 0.79 0.79 79%
7 LDA 0.65 0.65 0.65 65%
8 LOGISTIC REGRESSION 0.69 0.69 0.69 65%
9 NEAREST CENTROID 0.58 0.58 0.57 58%
10 PASSIVE AGGRESSIVE 0.69 0.69 0.69 69%
11 PERCEPTRON 0.68 0.67 0.67 67%
12 RIDGE 0.66 0.66 0.66 66%
13 SGD 0.68 0.68 0.68 68%
14 CNN 0.86 0.85 0.85 85%

16. CONCLUSION
• KNN AND THE DEEP LEARNING MODEL CNN PROVIDE BEST
ACCURACY FOR LUNG CANCER DETECTION FROM CT IMAGES
• THESE RESULTS CAN BE IMPROVED BY USING A BIGGER DATASET
OR OTHER PERFORMANCE ENHANCING TECHNIQUES

17. TESTING A CT IMAGE FOR LUNG CANCER
• I DEVELOPED A SIMPLE APPLICATION FOR CANCER DETECTION
USINT TKINTER LIBRARY OF PYTHON
• THIS APPLICATION ALLOWS USER TO SELECT ANY ALGORITHM
AND SELECT AN IMAGE TO GET THE RESULTS ABOUT THE IMAGE
• THIS ALSO SHOWS THE CONFUSION MATRIX AND
CLASSIFICATION REPORT OF THE DELECTED MODEL

18. APPLICATION

19. APPLICATION

20. APPLICATION

21. APPLICATION

22. APPLICATION

23. THE END