Accepted: EANM2022
Aim/Introduction: Coronary artery disease (CAD), also called ischemic heart disease, is the leading cause of death worldwide. Early CAD
detection is unavoidably critical for the patient’s health and, more specifically, for the determination of an effective treatment. To enable
reliable and trustworthy decisions, nuclear imaging specialists would require an autonomous diagnostic tool that can deliver accurate and at
the same time interpretable solutions. The proposed research deals with the development of an explainable machine learning pipeline that
is capable of identifying patients’ CAD status (normal, ischemia or infarction) through the use of a handcrafted Convolutional Neural
Network (RGB-CNN). The RGB-CNN network is accompanied with several pre- and post-processing tools including a state-of-the-art
explainability technique for understanding and visualizing the decisions behind CNN’s predictions. Materials and Methods: This project
aims to develop an explainable CNN-based approach that is sufficient in terms of scalability, robustness, and precision, while also being
capable of correctly differentiating SPECT-MPI heart images. The study includes a total of 630 patients in stress and rest representations.
Among them, 257 correspond to normal, 241 to ischemia and 127 to infarction, being originally classified by a doctor. The collected imaging
dataset was initially shuffled and then split into 15% for testing and 85% for training. The 15% of the training dataset was further used for
validation purposes. To achieve generability, data augmentation was used to introduce colour variations to the images. The best model was
concluded after a thorough exploration with numerous parameters (e.g. batch size, image size, number of nodes, number of convolutional
and dense layers), while each architecture was executed 10 times. Despite the adequate findings extracted by CNNs, they lack any rationale
for their decision-making. As a consequence, Grad-CAM was utilized in this research work to provide predictions with interpretability.
Results: The proposed model achieved 94.06% testing accuracy, 0.18 testing loss and 0.95% AUC, demonstrating great applicability to the
corresponding dataset and sufficient stability. Grad-CAM methodology was proved to be an effective tool in providing explanations of CNN-
based decision in SPECT MPI images, and as a result it could be adοpted as a supplementary post-hoc tool to support trustworthy clinical
diagnosis. Conclusion: Our proposed model achieved acceptable results, attaining high accuracy and minimized loss. This study contributes
to the effective diagnosis of ischemia and infarction in CAD, hence fostering trust in the use of explainable artificial intelligence models for
diagnosis in nuclear medicine.
