We are inspired by the Artificial Intelligence due to its extreme versatility in solving complex medical tasks.

What it does

Helping radiologists to overcome their high workload by automatically quantifying medical parameters from CT images of vertebrae to assess their fracture status.

How we built it

The fracture status can be defined by a semi-quantitative (SQ) classification proposed by Harry Genant (Genant et al, (1993) Vertebral fracture assessment using a SQ technique. J Bone Miner Res; 8:1137-48) with a rating from 0 to 4 towards more severe fracture status.

We trained a CNN model (DenseNet121) with 1000 labeled input images to do this classification.

Challenges we ran into

Few training images, especially the ratio of severe fractured cases versus Genant score 0. Import 16bit PNG images with varying histogram.

Accomplishments that we're proud of

We managed to compose a running CNN with classification and regression estimation of the Genant Score -
and of course we are proud of the things we learned and the good team spirit!

What we learned

Read the readme first :-) OR Do not write the export routine a few minutes before the deadline ;-)

What's next for "Backfisch"

  • Using a regression CNN to overcome „class imbalance“
  • Using another mapping to include the order of the different scores and train with a combination of binary cross-entropy loss and sigmoid activation function

Built With

  • imagej
  • keras
  • pycharm
  • python
  • tensorflow
  • vscode
Share this project: