Explainable CNN-Based Cardiac Amyloidosis Classification from PET Images Through Manifold Learning

Cardiac Amyloidosis (CA) is a myocardial disease characterised by the infiltration of misfolded proteins in the heart. Therapy varies according to the specific CA subtypes, and a good prognosis strongly depends on its accurate identification. However, CA can be easily misdiagnosed, especially in its early stage. Several imaging methodologies have been recently proposed to effectively and non-invasively detect and characterize CA, including PET imaging with [18F]-Florbetaben contrast. Convolutional Neural Networks (CNNs) have been reporting interesting performance in medical imaging classification tasks. However, their hardly interpretable architectures make it difficult to understand how they base their predictions. In dealing with high-stake decisions, it is important to overcome the black-box nature of such systems, but to date, there is still no comprehensive solution to the problem. One possible approach is to visualize in a reduced space through dimensionality reduction (DR) how a CNN processes input data to return its predictions. At present, several algorithms have been proposed, but to provide meaningful, non-misleading representations, it is crucial to assess such techniques with quantitative metrics. In our work, we applied different manifold learning algorithms (t-SNE, UMAP, TriMAP, and PaCMAP), subtypes of DR, to deal with the “open the black-box” problem, focusing on the case study of the diagnosis of Cardiac Amyloidosis (CA) from 2D [18F]-Florbetaben PET images. Our findings confirm the ability of PaCMAP to preserve both local and global data structures. In addition, quantitative metrics suggest that all the different manifold techniques are able to effectively map and visualize in a reduced space the features extracted by the CNN classifiers from the selected PET images.