Peeking inside the box : Transfer Learning vs 3D convolutional neural networks applied in neurodegenerative diseases

• Convolutional Neural Networks (CNNs) have shown their effectiveness in a variety of imaging applications including medical imaging diagnostics. However, these deep learning models are data-hungry and need enough labeled samples for the training phase which is limited in the medical domain. Transfer learning is one possible solution to this challenge with training a new model. Assessing model performance should be done not only based on criteria like accuracy, and area under the ROC curve, but also it is important to investigate what regions were of most interest for the classification decisions, especially for medical applications. We performed a case study on neurodegenerative disorders, in specific Alzheimer’s disease, mild cognitive im- pairment, dementia with lewy bodies and cognitively normal brains using 3D 18F-FDG-PET brain scans. Two transfer learning models, InceptionV3 and ResNet50, as well as a custom 3D-CNN that is trained from scratch are compared. Two XAI methods, occlusion and Grad-CAM are chosen to visualize the important brain regions using correctly classified cases. We found that the TL models learn significantly different decision surfaces than the 3D-CNN model. The 3D spatial structure of the brain regions are better kept in the 3D-CNN model, and that might explain the higher performance of this model over 2D-TL models. Moreover, we found out the two XAI methods provide different results, where occlusion method focused more on specific brain regions.

Ämnesord

TEKNIK OCH TEKNOLOGIER  -- Elektroteknik och elektronik -- Annan elektroteknik och elektronik (hsv//swe)

ENGINEERING AND TECHNOLOGY  -- Electrical Engineering, Electronic Engineering, Information Engineering -- Other Electrical Engineering, Electronic Engineering, Information Engineering (hsv//eng)

Nyckelord

Explainable artificial intelligence

neurodegenerative disease

transfer learning

occlusion

Grad-CAM

Publikations- och innehållstyp

ref (ämneskategori)

kon (ämneskategori)