An intrinsically explainable pipeline for MRI classification

Zenodo (CERN European Organization for Nuclear Research)

Abstract

AbstractMachine Learning systems achieve high diagnostic accuracy in medical imaging but their intransparent reasoning limits perceived trust and conflicts with regulatory requirements. While widely-used post-hoc explainable AI techniques only approximate the model’s reasoning, this thesis focuses on intrinsically explainable models, whose explanations are correct by design.An end-to-end pipeline for multi-class classification of 3D brain MRI images was developed. The pipeline was evaluated on Alzheimer’s Disease classification, but designed to generalize to other neurological diseases. The pipeline segments the brain into regions of interest using FastSurfer and trains a separate prototype-based PIPNet3D model per region. The prototype similarity values are combined in an interpretable ensemble classifier using logistic regression. Additionally, a neuro-symbolic classification approach was explored using inductive logic programming. Finally, an interactive web-based visualization tool enables clinicians to inspect the explanations. The proposed pipeline achieves comparable performance to whole-brain PIPNet3D, using only three regions with a F1-score of 82.8 ± 6.7 for binary classification and 47.9 ± 7.0 on multi-class classification. Discretizing prototype similarity values improves performance and enhances interpretability. Performance remains below black box approaches, though this is attributable to the feature extraction process. Explanations are evaluated using the Co-12 framework through functional and human-based assessment. By using an intrinsically explainable model, several properties are guaranteed by design. However, the human-based evaluation suggests prototypes only partially align with domain experts judgement.Future work includes a more comprehensive usability evaluation, integration of multi-modal imaging data and evaluating the pipeline on other neurological diseases.

Authors 1

1. Lars Quakulinski corresponding Aachen

   RWTH Aachen University

   Affiliation as printed

   RWTH Aachen University