Skip to content

Latest commit

 

History

History
54 lines (39 loc) · 2.39 KB

report.md

File metadata and controls

54 lines (39 loc) · 2.39 KB

Description of the implemented algorithm.

The CHOWDER algorithm falls under the scope of weak supervised learning. A first objective is to predict a binary label (Healthy VS Cancer) out of a given image. It is designed to operate over WSI images, that have a very high resolution. While the data is annotated at the image level, the network is fed with tiles obtained after sampling the original images. The algorithm exploits tiles descriptors to predict a label at the image level.

The algorithm can also be used to predict a label at the tile level (segmentation), a usage could be for example to help a clinician to annotate a dataset by suggesting suspicious regions on a WSI image.

Design choices and specifications of the code.

Code specs

The proposed application shall:

  • Obtain a path to medical data from the user, and load it in memory from disk.
  • train a classifier with CHOWDER method with training data and labels. The metric used to assess the performance of the classifier should be the AUC.
  • output evaluation results over testing data as a csv file called "test_output.csv"

The proposed application should:

  • log in console, and store on disk training and evaluation insights such as intermediate performance.
  • perform early stopping to keep the best trained model and store it on disk.

Design choices

The following design choices have been made, essentially to save time:

  • The slide data is zero padded, to handle the variation of the number of tiles, so that the batches have identical shapes
  • The CHOWDER method has been implemented without dropout and l2 regularisation.
  • The config parameters are written directly in the main script, they should ideally be written in a proper config file (in yaml for example)

Experimental results.

Model AUC score
Baseline+maxpool 0.65
CHOWDER (R=5) 0.857
CHOWDER (R=10) 0.837

Suggestions of improvement.

  • Add some visualisation routines (the ROC curve, but also the data during training at different steps)
  • Add a proper config file
  • Try the 3 fold cross-validation
  • Improve the model implementation by using drop out, l2 regularisation, changing activation function(?)
  • Test on another dataset to assess the robustness
  • More hyperparameters optimisation(grid search or random search)
  • Add more unit tests