1 of 31

Group 10

AI in Medical Imaging

​

1

Final presentation

楊佳誠 周軒民 陳品樺 鄭以新

​

​

2 of 31

2

  1. Motivation & Objective
  2. Methodology
  3. Result
  4. Conclusion

Outline

3 of 31

Motivation & Objective

Part 01

3

4 of 31

  • Health checks are about nipping any problems in the bud. The earlier you discover a problem, the earlier you can treat it.
  • False-negative test results result in delayed or lack of supportive treatment
  • Reach 90% or more accuracy in our model

​

​

4

4

Motivation & Objective – in our previous presentation

5 of 31

Methodology

Part 02

5

6 of 31

6

6

Methodology

Data preprocessing

Modify CNN code

Train our model

Improve our model

Prepare final presentation

7 of 31

7

7

Preprocessing

  • Resize the image to accelerate train speed

​

  • Revise defective data

​

  • Generate .csv to assist train

​

​

​

8 of 31

8

8

Preprocessing

​

    • Train data: 89.2%
    • Valid data: 0.2 %
    • Test data: 10.6%

​

9 of 31

9

9

CNN model

10 of 31

Result

Part 04

10

11 of 31

Model A :�First Attempt

12 of 31

12

12

Result_A

C

13 of 31

13

13

Result_A

14 of 31

14

14

Result_A

15 of 31

Model B: Add Gaussian Noise

16 of 31

16

16

Result_B (noise)

17 of 31

17

17

Result_B (noise)

18 of 31

18

18

Result_B (noise)

19 of 31

Model C: �Apply pre-trained initial weight

20 of 31

20

20

Result_C (pretrained initial weight)

21 of 31

21

21

Result_C (pretrained initial weight)

22 of 31

22

22

Result_C (pretrained initial weight)

23 of 31

Model D: �Inception-V3

24 of 31

D model visualization

25 of 31

25

25

Result_D (Other model)

26 of 31

26

26

Result_D (Other model)

27 of 31

27

27

Result_D (Other model)

28 of 31

29 of 31

Conclusion

Part 05

29

30 of 31

30

30

Conclusion

  • Prevent overfitting
    • Add Gaussian noise

​

  • Improve performance
    • Initial weight

​

  • With well-trained model, it can support doctor’s judgments

​

​

31 of 31

Thanks For Listening

31