Deep learning techniques are increasingly utilized in medical applications, including image reconstruction, segmentation, and classification. However, despite the good performance those models are not easily interpretable by humans. Especially medical applications require to verify that this is not the result of exploiting data artifacts.

Our experiments on Alzheimer's disease (AD) classification showed that Deep Neural Networks (DNN) might learn from features introduced by the skull stripping algorithm. Therefore, we are investigating how preprocessing (registration and brain extraction) determine which and how many features in the MR images are relevant for the separation of patients from healthy controls.

Relevance-Guided Deep Learning (Graz⁺)

We develop a regularization technique to train convolutional neural network (CNN) classifiers utilizing relevance-guided heat maps calculated online during training. The developed relevance-guided framework achieves higher classification accuracies than conventional CNNs but more importantly, it relies on less but more relevant and physiological plausible voxels within brain tissue. Additionally, preprocessing effects from skull stripping and registration are mitigated, rendering this practically useful in deep learning neuroimaging studies.

Mean heat maps (highest relevances in yellow, overlaid on MNI152 template) and balanced classification accuracy (percentage). Unmasked and masked CNN classifiers obtain relevant image features overwhelmingly from global volumetric information (left and center columns), whereas Graz⁺ exclusively relies on  deep gray and white matter tissue adjacent to the ventricles (right column).

The preprint paper can be found below and the source code for the Graz⁺ technique is available under https://github.com/christiantinauer/Graz_plus_technique.

References

Tinauer C, Heber S, Pirpamer L, Damulina A, Schmidt R, Stollberger R, Ropele, S, Langkammer C.
Interpretable brain disease classification and relevance-guided deep learning.
Sci Rep. 2022 Nov 24;12(1):20254. doi:10.1038/s41598-022-24541-7. PMID: 36424437; PMCID: PMC9691637.

Abstracts

C Tinauer et al., Increasing Feature Sparsity in Alzheimer's Disease Classification with Relevance-Guided Deep Learning
ISMRM, 2021

C Tinauer et al., Relevance-guided Deep Learning for Feature Identification in R2* Maps in Alzheimer’s Disease Classification
ISMRM, 2020

C Tinauer et al., Relevance-guided Feature Extraction for Alzheimer's Disease Classification
ISMRM, 2019

Heatmapping

Wonderful resource about explainability in deep learning: http://heatmapping.org.