Investigating methods to enhance interpretability and performance in cardiac MRI for myocardial scarring diagnosis using convolutional neural network classification and One Match

Abstract

Machine learning (ML) classification of myocardial scarring in cardiac MRI is often hindered by limited explainability, particularly with convolutional neural networks (CNNs). To address this, we developed One Match (OM), an algorithm that builds on template matching to improve on both the explainability and performance of ML myocardial scaring classification. By incorporating OM, we aim to foster trust in AI models for medical diagnostics and demonstrate that improved interpretability does not have to compromise classification accuracy. Using a cardiac MRI dataset from 279 patients, this study evaluates One Match, which classifies myocardial scarring in images by matching each image to a set of labeled template images. It uses the highest correlation score from these matches for classification and is compared to a traditional sequential CNN. Enhancements such as autodidactic enhancement (AE) and patient-level classifications (PLCs) were applied to improve the predictive accuracy of both methods. Results are reported as follows: accuracy, sensitivity, specificity, precision, and F1-score. The highest classification performance was observed with the OM algorithm when enhanced by both AE and PLCs, 95.3% accuracy, 92.3% sensitivity, 96.7% specificity, 92.3% precision, and 92.3% F1-score, marking a significant improvement over the base configurations. AE alone had a positive impact on OM increasing accuracy from 89.0% to 93.2%, but decreased the accuracy of the CNN from 85.3% to 82.9%. In contrast, PLCs improved accuracy for both the CNN and OM, raising the CNN's accuracy by 4.2% and OM's by 7.4%. This study demonstrates the effectiveness of OM in classifying myocardial scars, particularly when enhanced with AE and PLCs. The interpretability of OM also enabled the examination of misclassifications, providing insights that could accelerate development and foster greater trust among clinical stakeholders.

Fig 1. One Match (OM) workflow.

A) Input image. B) The input image is matched with the positive and negative template sets, finding the highest correlation for each set. C) The highest correlation from each template set is then compared. D) The input image is classified based on which template set has the higher correlation. In this case, the negative template set has a higher correlation, so the classification is negative.

Fig 2. Architecture and workflow for sequential convolutional neural network (CNN).

Lightweight Preprocessing was utilized to transform cardiac MRI data into the Input. The input data was then passed through the CNN to train it and make predictions about the presence of myocardial scarring in the images.

Fig 3. Autodidactic enhancement workflow.

A) The true label of each input image is known. B) Each input is matched against the images in the positive and negative template sets, checking to see if it matches to the correct images and doesn’t match with the wrong ones. C) The number of correct and incorrect matches are summed. D) The decision is a simple majority decision. When the number of correct images is greater, the image is retained. Otherwise, the image is discarded.

Fig 4. Classification workflow comparison for the sequential convolutional neural network (CNN) and One Match (OM).

Input was first processed by Lightweight Preprocessing (LWP) and optionally processed with autodidactic enhancement (AE). Classifications were then performed with either the sequential CNN or with OM. Patient-level classifications (PLCs) were applied optionally as well.

Fig 5. Confusion matrices for myocardial scarring (MS) classification with patient-level classifications.

A) Cross-validated results for the convolutional neural network (CNN) without the autodidactic enhancement (AE). B) Cross-validated results for the CNN with the AE. C) Results for One Match (OM) without the AE. D) Results for OM with the AE. For the CNN, AE reduced both sensitivity and specificity, while it increased sensitivity for OM.

Fig 6. Image misclassified by One Match (OM).

A) Misclassified image with scarring indicated by red oval. B) Template set image that had the highest correlation. C) Blend of part A and part B, cyan indicates regions present in A but not in B, and magenta indicates regions present in B but not in A. The area of scarring, highlighted by a red oval, appears in cyan. Even with this scar present, the 0.841 correlation between part A and part B resulted in part A being classified as not having scar.

Fig 7. Image without scarring misclassified by One Match (OM).

A) Misclassified image without myocardial scarring. B) Template image that had the highest correlation. C) Blend of part A and part B, with cyan indicating which regions present in A not present in B and magenta indicating the opposite. Despite what might be considered drastic differences by a human clinical observer, these two images yielded a correlation score of 0.826.