Gustave Roussy

Data Challenge
JFR 2018

Several multidisciplinary teams working on the development of AI algorithms during the Journées Francophones de Radiologie

COMPUTER VISION
AI
HEALTH

Context

For the first time, the Journées Francophones de Radiologie, in partnership with the Gustave Roussy Institute, are hosting an AI Forum and have set up a unique data challenge in the medical field.

We worked, in collaboration with radiologists, on the detection of meniscal lesions from knee MRIs.

We achieved a performance of 84.6% on a hybrid metric comprised of lesion detection, orientation and location.

SOLUTION

Sobel Filter, data augmentation, and CNN.

Image preprocessing

• Data enrichment with a Sobel filter:
We extracted the
images’ contours, adding valuable information to our dataset.

• Image normalization: frame by frame (samplewise)

• Data augmentation: to increase the number of images and allow for better generalization, we have applied:
o Random rotation of the images (maximum +/- 45 degrees)
o Horizontal random translation (between +/- 20% of the length).
o Vertical random translation (between +/- 20% of the length)

.

Model and Training

After multiple iterations with various models (Transfer Learning, Autoencoder, etc…), we finally adopted the following architecturet:

  • • 3 distinct binary classification models, trained separately on the 3 tasks
  • • 3 or 4 convolutional layers trained “from scratch” (kernel: 7, 5, 3, 3, 3; filters: 64, 32, 16, 8)
  • • 3 dense layers: 32, 16, 1
  • • light dropout for convolutional layers

The training is carried out with classic parameters: optimizer=Adam, lr=0.001 (reduced on the plateaus), batch_size=32, epochs<=200, loss=binary_cross_entropy

Inference

The image prediction process differs from the training process simply by how the data augmentation is performed. The images are augmented by the same process, but
with lower values for rotation (+/- 20 degrees), in order to resemble the raw images. We empirically observed that this intuition allowed for better
predictions. 5 new images are generated, and the predictions on the 6 modalities (raw image + 5 augmented images) are averaged for the final prediction.

Benefits

A first participation in a medical imaging competition.

Experience

Thanks to our computer vision skills, this competition allowed us to challenge ourselves in a new field: medical imaging, and at the same time extend our internal Machine Learning library.

Competition

Having finished 2nd overall out of 11 teams without the necessary help on image segmentation, a simpler model still allowed us to be only 6% away from the best team (90.6%). We look forward to the next challenge!

FOCUS

Some visuals of our approach.