Informed Machine Learning for Cardiomegaly Detection in Chest X-Rays - A Comparative Study

Aug 5, 2026

Upload by Aurora

Informed machine learning (IML) is applied to cardiomegaly detection from chest radiographs to address practical limits of convolutional neural networks, including insufficient data quality and limited data availability. The study implements, combines, and evaluates three IML approaches against a purely data-driven baseline on the CheXpert dataset. Results show that curriculum learning and cropping images to regions of interest improve prediction performance. The work provides reference guidance for implementing, evaluating, and combining IML methods for future research.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages5

Reading Time13 min

AccessFree to download

Introduction

Background

INFORMED MACHINE LEARNING FOR CARDIOMEGALY DETECTION IN CHEST X-RAYS: A COMPARATIVE STUDY Felix Hasse, Florian Leiser, Ali Sunyaev Karlsruhe Institute of Technology Department of Economics and Management ABSTRACT Recently, computer-aided disease detection from chest radiographs made considerable progress by using convolutional neural networks but issues like insufficient data quality or data availability remain. Informed machine learning (IML) combines domain knowledge and data-driven approaches and has been shown to improve results in many applications. However, there is limited research comparing and combining multiple IML approaches. This paper tackles this issue by implementing, combining, and evaluating three IML approaches for cardiomegaly detection. We find that curriculum learning and cropping images to regions of interest can improve prediction performance. With these results, we provide a reference for both implementing and evaluating multiple IML approaches as well as demonstrating methods to combine IML approaches. Index Terms—informed machine learning, cardiomegaly, CheXpert, curriculum learning 1. INTRODUCTION Cardiomegaly, an abnormal enlargement of the heart, is a frequent indicator for other pathologies, like coronary artery disease, myocardial infarction, and ischemia [1] . Previously, clinicians investigated X-rays manually to identify these conditions. Machine learning (ML) and especially convolutional neural networks (CNNs) allowed for considerable progress in early and accurate disease detection from chest X-rays over the last years [2, 3] . However, challenges remain when using CNNs in practice. In particular, the models often perform badly if training data is limited [4] . Informed ML (IML) addresses this and other issues by incorporating external prior knowledge into ML models. Currently, however, practitioners do not know which IML approaches are best suited for their task [5] . Previous research implementing IML approaches only evaluated the approaches individually or compared them to a purely data-driven baseline [4] . Due to differences in knowledge, data, or model architectures [6], it is difficult to compare results across papers. This motivates the need for a detailed comparison between approaches on the same dataset using the same model architecture. Additionally, since different types of knowledge are incorporated into IML models, leveraging different approaches by combination could further improve performance. Therefore, we pose the following research question: How do different IML approaches and their combination compare to each other? To answer this question, we implement three different IML approaches and a purely data-driven baseline model and evaluate them on the CheXpert dataset [3] . The first approach is based on calculating the cardiothoracic ratio (CTR) . CTRis the ratio of the heart diameter to the internal chest diameter on a posterior-anterior radiograph or computed tomography. Cardiomegaly is present when CTR is ≥ 0.5 [1, 7] . Here, the heart and lung diameters are segmented from X-ray images to predict an abnormal heart-to-lung diameter ratio. The second approach is curriculum learning where models are trained on easier examples first [8] . We use the predicted CTR as a proxy to score the difficulty of training samples. Our third approach investigates only the image region relevant to cardiomegaly diagnosis. To achieve that, we train a CNN on images cropped to bounding boxes extracted by generated lung segmentation masks. Finally, we implement a model combining curriculum learning and image cropping. Afterward, we evaluate all IML approaches’ predictive performance for cardiomegaly detection on varying training dataset sizes. With this study, we demonstrate new ways to combine IML approaches and provide meaningful comparisons between approaches allowing for easier future comparison. Furthermore, the used methods may provide a reference for future res

FAQs

What problem does informed machine learning aim to solve for cardiomegaly detection from chest X-rays?

It addresses issues seen in practice when CNNs have limited or insufficient data quality and availability by incorporating external prior domain knowledge into the learning process.

Which three IML approaches are implemented in the study, and what does each do?

The first uses CTR calculation from segmented heart and lung regions; the second applies curriculum learning by ordering samples using predicted CTR as a proxy for difficulty; the third crops images to regions relevant to cardiomegaly using lung segmentation masks, and a combined approach uses both curriculum learning and cropping.

How does the study evaluate performance across different training dataset sizes?

All approaches, including the data-driven baseline, are evaluated on the CheXpert dataset while varying the amount of training data to compare predictive performance under different data availability conditions.

Aug 5, 2026

Upload by Aurora