Weakly Supervised Learning for
Findings Detection in Medical Images
HaoCheng Kao
Outline
• The Task
• A Naïve method
• The Dataset
• Grading
• Submission
• Rules
• Contacts
Task
Task
Task
• 14 Findings (classes)
• Predict bounding box + class
• Training data has only class labels
Naïve method
Dataset
• We use NIH ChestX-ray 14 dataset
https://nihcc.app.box.com/v/ChestXray-NIHCC
• Training set (111240)
• Validation set (440)
• Testing set (440)
Download
• Data_Entry_2017_v2.csv
Image-level annotation for all images.
• Bbox_List_2017.csv
Bounding box annotation for validation / testing images.
• train.txt / valid.txt / test.txt Lists of images in each set.
Grading
• 60% accuracy
• 40% novelty & report or poster
Intersection over Union
Accuracy
• For each ground truth box in a test image, we’ll check if there is an output box correctly locate this box.
– Class label is correctly predicted.
– IoU of ground truth box and predicted box >= T(IoU).
• The ratio of successful ground truth boxes in an image is the score of the image.
• The average score of all images is the score of a specific T(IoU).
• The final score is the average of T(IoU)=0.25 and T(IoU)=0.5.
• You must output up to 10 bounding boxes per image.
Typical workflow of a submission
1. Call get_file_names() to get the list of files.
2. Load your model.
3. For each image, inference the answer.
4. Call get_output_file_object() to get the output file object.
5. Write the output.
6. Call judge() to get the result.
Note that the timestamp created when judge() is called is used for judging if the submission is on time.
Submission
• Submission for evaluating on testing data
– During this competition, you can submit your model to be evaluated on testing data.
• Final submission
– After you complete your project, you should
submit your whole project so that we can verify your result.
Final Submission
• Your final submission should contain the following:
– Trained model & Whole Code – Document & report
Task Restrictions
• Keras / PyTorch / TensorFlow / MXNet / CNTK
• You can only submit a single archive
containing all the codes and model. This archive must not exceed 1GB in size.
Restrictions on Dataset
• Stick to the dataset provided by link.
• The bounding boxes of validation / testing set are publicly disclosed, but you may not use
them to train your model.
• You are allowed to use only the following extra context:
Patient Age / Patient Gender / View Position
Pre-trained Model
• The pre-trained model should only be trained by the ImageNet dataset for classification and
localization.
• You can download any available pre-trained model from the internet
– Make sure that it is only trained by the ImageNet dataset for classification and localization.
– Make sure you have the rights to use the model.
• You can create your own pre-trained model, but it should be trained from randomly-initialized
parameters.
Rules
• The final model must only rely on the training / validation set. The bounding boxes of the validation set can only be used to validate your model.
• For grading, you must submit your model / algorithm to the submission site before the deadline following the
submission rules.
• All submitted materials must be created by the team
members or the team members must have proper rights to use them.
• All novelty must be created solely by the team members.
No assistance from outside of the team is allowed.
• Course regulations, the contract, and school regulations bust be followed.
Contacts
• HaoCheng_Kao@htc.com
• In general, we will announce your questions and our answers to all students instead of replying to each email.
