Leveraging Explainability methods in Spectral Domain for Data Augmentation and efficient training of CNN classifiers for Covid-19 Detection

In our current work we tackle the complex problem of classifying medical images in X-ray modality to distinguish between lungs with Covid-19 and Normal control subjects. We aim to use explainability methods to identify specific frequencies in the spectrum which are responsible for the contrasts between Covid "Clouds" and surrounding lung and bowl tissues. Hence, we first trained a ResNet-50 deep neural network classifier on the frequency domain representation and then applied previously developed Feature based Explanation Method (FEM) explainer to this spectrum. The explanation mask in spectral domain is then used as an importance filter. We then propose a data augmentation technique where instead of using conventional label preserving filtering such as a weak low pass Gaussian filter, we use this importance mask synthesised in the frequency domain to emphasize the frequencies that contributed to the network decision and attenuate the non-important ones. A specific training strategy, altering between original image and importance filtered reconstructed image allowed for an increase of performance of binary classifier up to 2% in terms of Balanced accuracy and 6% in its sensitivity.

Meaza Eyakem Gebreamlak, Meghna Ayyar, Jenny Benois-Pineau, Jean-Pierre Salmon, Akka Zemmari. Leveraging Explainability methods in Spectral Domain for Data Augmentation and efficient training of CNN classifiers for Covid-19 Detection. 2023 Twelfth International Conference on Image Processing Theory, Tools and Applications (IPTA), Oct 2023, Paris, France. pp.1-6, ⟨10.1109/IPTA59101.2023.10320061⟩ - lien externe. ⟨hal-04684288⟩ - lien externe