Encoding of Demographic and Anatomical Information in Chest X-Ray-Based Severe Left Ventricular Hypertrophy...

Area of Science:

• Artificial Intelligence in Medicine
• Cardiology
• Medical Imaging Analysis

Background:

• Severe left ventricular hypertrophy (SLVH) is a significant risk factor for heart failure.
• Current diagnostic methods like echocardiography and MRI are costly and burdensome.
• There is a need for accessible and efficient SLVH detection methods.

Purpose of the Study:

• To develop and validate a deep learning framework for SLVH classification directly from chest radiographs.
• To evaluate the interpretability of the deep learning model by quantifying attribute encoding.
• To assess the potential of AI in improving cardiac abnormality detection workflows.

Main Methods:

• Utilized a class-balanced subset of the CheXchoNet dataset.
• Fine-tuned a ResNet-18 model and pretrained a Vision Transformer (ViT) encoder.
• Applied Mutual Information Neural Estimation (MINE) to analyze feature interpretability regarding clinical attributes.

Main Results:

• The Vision Transformer (ViT) model achieved an AUROC of 0.82 and AUPRC of 0.80 for SLVH detection.
• MINE analysis demonstrated that the model effectively encodes clinical attributes like age, sex, and cardiac dimensions.
• The model showed strong performance without requiring demographic or intermediate anatomical data.

Conclusions:

• Chest radiographs alone are sufficient for accurate SLVH classification using deep learning.
• The developed framework offers both diagnostic accuracy and quantitative interpretability.
• This AI approach shows promise for enhancing clinical decision support and patient triage systems.