OC 24.5 - Leveraging Explainable Artificial Intelligence Methods for Hemophilic Arthropathy: A Potential Paradigm Shift in Diagnosis and Management

Background: Despite the apparent potential of deep learning to augment Petterson score assignment in radiographic hemophilia arthropathy evaluations, a significant research lacuna exists regarding the core factors influencing these models' judgment process.

Aims: This study aims to leverage the power of heatmaps as a visual exploration tool to delve into the key regions of hemophilia knee arthropathy, shedding light on the distribution and intensity of joint damage.

Methods: Leveraging a curated dataset of 438 knee radiographs from 44 hemophiliac patients acquired between 2016 and 2022, this study explores the utility of heatmaps in visualizing "hot spots" associated with radiographic hemophilia knee arthropathy. We employ a machine learning model trained on meticulously labeled data to predict disease severity, then utilize explainable AI techniques to generate heatmaps highlighting the anatomical regions influencing model predictions.

Results: In this study, EfficientNet-b5 was used to train the model with an accuracy of 94.98%, and the ROC Curve showed excellent results. Heatmaps unveiled the location of "hotspots" directly linked to hemophilia knee arthropathy's severity.

Conclusion(s): This study demonstrates the effectiveness of utilizing both an EfficientNet-b5 model and heatmap analysis for assessing hemophilic knee arthropathy. The high accuracy, combined with the visualization of disease hotspots, presents promising avenues for advancing clinical decision-making and potentially tailoring treatment strategies for optimal patient outcomes.