Deep learning to detect left ventricular structural abnormalities in chest X-rays

Shreyas Bhave, Victor Rodriguez, Timothy J. Poterucha, Simukayi Mutasa, Dwight Aberle, Kathleen M. Capaccione, Yibo Chen, Belinda D’Souza, Shifali Dumeer, Jonathan Goldstein, Aaron Hodes, Jay Leb, Matthew P. Lungren, Mitchell D. Miller, David Monoky, Benjamin Navot, Kapil Wattamwar, Anoop Wattamwar, Kevin J. Clerkin, David Ouyang, Euan A. Ashley, Veli K. Topkara, Mathew Maurer, Andrew J. Einstein, Nir Uriel, Shunichi Homma, Allan Schwartz, Diego Jaramillo, Adler Perotte, Pierre EliasEuropean 2024 · DOI

TL;DR

This paper introduces a deep learning model that can accurately detect signs of heart failure, specifically severe left ventricular hypertrophy (SLVH) and dilated left ventricle (DLV), from standard chest X-rays (CXRs). Remarkably, the model not only matches but surpasses the performance of board-certified radiologists in this task, offering a scalable and cost-effective screening tool for early heart failure detection.

What Problem Does This Paper Address?

Early identification of heart failure is crucial for improving patient outcomes, but current diagnostic methods like echocardiography are often reserved for patients with a high pre-test probability. This leads to delays in diagnosis, especially for those with early-stage heart failure who may not yet exhibit overt symptoms. Chest X-rays (CXRs) are a ubiquitous and inexpensive imaging modality, making them an ideal candidate for a broader screening tool. However, traditional CXR interpretation for subtle structural abnormalities indicative of heart failure has been limited. Previous deep learning efforts have primarily focused on more general findings like cardiomegaly, which are not specific enough for actionable clinical decisions. This paper aims to bridge this gap by developing and validating a deep learning model that can reliably detect specific structural abnormalities (SLVH and DLV) from CXRs, which are key indicators of heart failure.

Methods

The researchers developed a deep learning model based on the DenseNet-121 architecture. The model takes a pre-processed chest X-ray image, along with patient age and sex, as input. It's trained to output probabilities for three conditions: severe left ventricular hypertrophy (SLVH), dilated left ventricle (DLV), and a composite label indicating the presence of either.

Key methodological aspects include:

  • Data Pre-processing: CXRs were cropped to a 1:1 aspect ratio and downsampled to 224x224 pixels. Contrast was enhanced using adaptive histogram equalization, and Gaussian noise was added to the training set to improve generalization.
  • Model Architecture: DenseNet-121 was chosen for its proven ability to learn effective representations from medical images.
  • Label Generation: Ground truth labels for SLVH and DLV were derived from echocardiogram measurements, using established clinical guidelines for thresholding.
  • Model Evaluation: Performance was assessed using Area Under the Receiver Operating Characteristic Curve (AUROC) and Area Under the Precision-Recall Curve (AUPRC). The model was evaluated on an internal test set, an external validation set, and compared against human radiologists.
  • Explainability: LayerCAM was used to generate saliency maps, visualizing which parts of the CXR the model focused on for its predictions, helping to ensure it was attending to relevant cardiac structures rather than spurious artifacts.

Datasets

  • Columbia University Irving Medical Center (CUIMC) Dataset: This was the primary dataset, comprising 71,589 unique CXRs from 24,689 patients, with echocardiogram labels obtained within 12 months of the CXR. This dataset was split into training (90%), validation (5%), and test (5%) sets.
  • Stanford University Medical Center Dataset: An external dataset of 8,003 CXRs from 4,657 patients, used for external validation.

The paper also makes the CUIMC dataset of 71,589 CXRs with echocardiographic labels publicly available to foster further research.

Key Insights and Ideas

  • CXRs Hold Significant Diagnostic Power: The study demonstrates that standard chest X-rays, often considered a less detailed imaging modality for cardiac assessment, contain sufficient information to detect specific structural abnormalities indicative of heart failure.
  • Deep Learning Outperforms Human Experts: A significant finding is that the developed deep learning model not only matches but surpasses the performance of board-certified radiologists in identifying SLVH and DLV from CXRs. This is a critical step towards AI-driven diagnostic assistance in cardiology.
  • Scalable Screening Potential: The ability of the model to perform well on CXRs, a widely available and inexpensive imaging modality, opens up possibilities for large-scale, cost-effective screening for early signs of heart failure, potentially identifying patients before they become symptomatic.
  • Model Focuses on Relevant Anatomy: Saliency maps indicate that the deep learning model attends to the cardiac silhouette and structures of the left ventricle, suggesting it's learning clinically relevant features rather than relying on artifacts or confounders.

Key Results

  • Model Performance (CUIMC Test Set):
    • SLVH: AUROC of 0.79 (95% CI 0.76-0.81)
    • DLV: AUROC of 0.80 (95% CI 0.77–0.84)
    • Composite (SLVH/DLV): AUROC of 0.80 (95% CI 0.78–0.83)
  • External Validation (Stanford Set): The model maintained reasonable performance, with an AUROC of 0.76 for the composite label, despite differences in data distribution.
  • Performance on Pre-Symptomatic Patients: When evaluated on CXRs taken before the first echocardiogram, the model achieved an AUROC of 0.80 for the composite label, indicating its utility in identifying at-risk individuals before a formal diagnosis.
  • Radiologist Comparison: The model outperformed all 15 individual radiologists on a subset of images. At a specificity of 73%, the model achieved a sensitivity of 71%, compared to the radiologists' consensus vote of 66%.

Limitations and Future Work

  • Data Distribution Shifts: The performance drop on the SLVH label in the external Stanford dataset highlights challenges with data distribution shifts (differences in imaging equipment, patient populations, or labeling practices). The authors suggest that a larger, multi-site study is needed to assess robustness.
  • Echocardiogram Label Variability: Echocardiographic measurements, while considered gold-standard, are subject to inter-observer variability. The study attempted to mitigate this by modeling it as a variance parameter.
  • Limited External Data: Validation was performed on only one external site, limiting the generalizability assessment.
  • Temporal Proximity of Labels: While the study used labels within 12 months, patients in the "pre-first echo" analysis might still be closer to a diagnosis than truly asymptomatic individuals.
  • Explainability Nuances: While saliency maps provide insights, deep learning models remain largely opaque in explaining individual predictions, posing a challenge for building user trust.

Future Work: The authors suggest that prospective validation of such a model is the necessary next step. They also propose exploring how radiologists can best interact with this AI technology in a clinical workflow, and leveraging CXRs as an ideal modality for capturing a proportion of the underdiagnosed heart failure population.

Summary updated

Abstract

Deep learning analysis of CXRs can accurately detect the presence of certain structural abnormalities and may be useful in early identification of patients with LV hypertrophy and dilation. As a resource to promote further innovation, 71 589 CXRs with adjoining echocardiographic labels have been made publicly available.
BibTeX
@inproceedings{bhave2024deep,
  author = {Shreyas Bhave and Victor Rodriguez and Timothy J. Poterucha and Simukayi Mutasa and Dwight Aberle and Kathleen M. Capaccione and Yibo Chen and Belinda D’Souza and Shifali Dumeer and Jonathan Goldstein and Aaron Hodes and Jay Leb and Matthew P. Lungren and Mitchell D. Miller and David Monoky and Benjamin Navot and Kapil Wattamwar and Anoop Wattamwar and Kevin J. Clerkin and David Ouyang and Euan A. Ashley and Veli K. Topkara and Mathew Maurer and Andrew J. Einstein and Nir Uriel and Shunichi Homma and Allan Schwartz and Diego Jaramillo and Adler Perotte and Pierre Elias},
  title = {{Deep learning to detect left ventricular structural abnormalities in chest X-rays}},
  booktitle = {European},
  year = {2024},
  doi = {10.1093/eurheartj/ehad782},
}
Figure