1 of 22

Machine Learning Final Project:

Medical Image Detection

MLTAs

ntumlta2019@gmail.com

2 of 22

Outline

  • Task Description - Medical Image Detection
  • Data Format
  • Kaggle
  • Requirements
  • FAQ

3 of 22

Outline

  • Task Description - Medical Image Detection
  • Data Format
  • Kaggle
  • Requirements
  • FAQ

4 of 22

Task Description

  • Lung disease detection

Normal

Diseased

5 of 22

Task Description

  • Lung disease detection

Ground truth

Predicted

6 of 22

Task Description

  • How to?

​

​

​

​

​

​

  • Hint: pretrained CNN may help!!

​

​

​

​

​

CNN feature extractor

DNN output layer for bounding box prediction

7 of 22

Task Description

  • Procedure
    • Resize pictures to the same size
    • Normalize the ground truth bbox with respect to the width/height of each picture
    • Train the network
    • Predict normalized bbox and un-normalize it
  • Loss
    • Binary classification: normal/diseased
    • Position of the bbox

​

8 of 22

Task Description: Evaluation Metrics

  • Intersection over union score
    • A metrics for image segmentation
    • Treat the detection problem as a segmentation problem by simply labeling the pixels within the bbox as 1, out of the bbox as 0

​

9 of 22

Task Description: You may be interested

​

10 of 22

Task Description: You may be interested

​

11 of 22

Outline

  • Task Description - Medical Image Detection
  • Data Format
  • Kaggle
  • Requirements
  • FAQ

12 of 22

Data Format

  • File layout
    • Data/ --- train_labels.csv

|--train_images/

|--test_images/

13 of 22

Data Format

  • Train_labels.csv
    • Each line stand for one bbox, instead of one picture!!!
    • PatientId: filename for the picture
    • x, y: the up-left corner of the bbox
    • width, height: the width and height of the bbox,� measured in pixels
    • Target: 1 for diseased, 0 for healthy

​

x

y

w

h

14 of 22

Data Format

  • Train_labels.csv

​

healthy, thus no bbox

multiple bboxes for train-00003.png

15 of 22

Outline

  • Task Description - Medical Image Detection
  • Data Format
  • Kaggle
  • Requirements
  • FAQ

16 of 22

Kaggle

​

17 of 22

Kaggle

  • Submission format: run-length encoding for pixel-wise segmentation masks

​

18 of 22

Kaggle

  • Label the pixels in the bboxes as 1, others as 0
  • Run-length encoding of the 1-pixels
    • The competition format requires a space delimited list of pairs
    • For example, '1 3 10 5' implies pixels 1,2,3,10,11,12,13,14 are to be included in the mask
    • The metric checks that the pairs are sorted, positive, and the decoded pixel values are not duplicated
    • The pixels are numbered from left to right, then top to bottom(e.g.1 is pixel (1,1), 2 is pixel (1,2), etc.)
  • Don’t worry, TA provides the code to transform the train_labels.csv format into the kaggle submission format!

​

19 of 22

Outline

  • Task Description - Medical Image Detection
  • Data Format
  • Kaggle
  • Requirements
  • FAQ

20 of 22

Requirements

  • Any method is allowed, excluding…
  • Use your classmate’s code
  • Use the labels of the test data directly or indirectly. (Do not try to find them.)
  • Train your model on any other dataset(but pretrained CNN is allowed)
  • Pretrain your CNN on dataset other than NIH-Chest X-ray dataset and ImageNet dataset
  • Submit prediction with more than one Kaggle account
  • Give/get model prediction to/from others
  • Give/get trained model to/from others
  • Publish your code before deadline

​

21 of 22

Outline

  • Task Description - Malicious Comments Identification
  • Data Format
  • Kaggle
  • Requirements
  • FAQ

22 of 22

FAQ

  • 若有其他問題，請寄信至助教信箱，請勿直接私訊助教。
  • 有問題建議可以在 FB Group 裡面留言發問，可能很多人都有一樣的問題
  • 不足之處請參照deepQ提供的投影片，關於kaggle以及競賽方面規定若有衝突以deepQ的投影片為主
  • 助教信箱: ntumlta2019@gmail.com
  • Useful Website: link

​