Model training

[solomesolo]

Training ~3 models on the dataset. It will be ~10 days with retraining models with little parameter tuning.

[bogdan-fesenko]

The upload of network training data to the Tensorboard was added to analyze the learning process in real-time. Kappa and ROC metrics were added to calculate the effectiveness of training a neural network on a validated dataset in real-time.

[bogdan-fesenko]

Trained densenet169 model and for now got the best performance with single view evaluation: Kappa score: 0.652 AUC ROC score : 0.830 Accuracy : 0.827 But for this competition, the prediction was evaluated for the entire study, and not a single image.

[bogdan-fesenko]

The goal is to achieve accuracy comparable to the results from Stanford's scientific work, which are 0.705 for Kappa score and 0.929 for AUROC. This can be achieved when we'll make the same prediction process, that is, add predictions using an ensemble, and not just one neural network and make a general prediction for the entire study, based on all views

[bogdan-fesenko]

Achieved when predicting for the whole study. 2.66 images per study on average. Kappa score: 0.700 AUC ROC score : 0.856 Accuracy : 0.853

[solomesolo]

Achieved when predicting for the whole study. 2.66 images per study on average. Kappa score: 0.700 AUC ROC score : 0.856 Accuracy : 0.853 @bogdan-fesenko Is this the result of our model?

[solomesolo]

0.830 @bogdan-fesenko Is it results when using a few networks simultaneously?

[bogdan-fesenko]

Yes, it's the result of our trained model with the prediction by a single model, but for a few views.