GradCAM as an explicability method to evaluate the performance of deep learning models in classifying pediatric arteriovenous malformations (AVM) in arterial spin labeling sequences (ASL)

Abstract

Purpose

The study investigates the usefulness of Convolutional Neural Networks (CNNs) in accurately detecting arteriovenous malformations in pediatric medical imaging, particularly using arterial spin labeling sequences. It also aims to offer diagnostic explanations comparable to expert analysis. Methods

The research analyzed three different CNN architectures to determine their performance in detecting arteriovenous malformations. The study focused on evaluating the relationship between model complexity and performance increase, using data to assess the accuracy and diagnostic usefulness of each model. Results

The findings indicated a nonlinear link between model complexity and performance. Sur- prisingly, more complex models frequently produced poor results and diagnostically useless answers. The simplest CNN models achieved the highest accuracy rate (90%), demonstrating the effectiveness of minimal complexity in model construction. Heat maps showed a strong association with the real locations of irregularities, indicating that the models were interpretable. Conclusion

The study highlights the usefulness of CNNs in medical diagnostics, emphasizing the importance of model simplicity and interpretability in clinical applications. It suggests a need for balancing technical sophistication with clinical value and presents options for future research into refining CNN structures for increased diagnostic precision in various medical imaging modalities.

Peer Reviewed

Postprint (author's final draft)