Automating Detection of Papilledema in Pediatric Fundus Images with Explainable Machine Learning

Kleanthis Avramidis, Mohammad Rostami, Melinda Chang, Shrikanth Narayanan

Abstract

Distinguishing between papilledema and pseudopapilledema in pediatric patients is a critical clinical challenge, as the former indicates life-threatening intracranial pressure while the latter is a benign anatomical variation. This paper presents a multi-branch deep learning framework designed to automate this classification using fundus images. By employing unsupervised optic disc localization, clinically-inspired color channel augmentations, and a combination of cross-entropy and triplet loss, the proposed model achieves an accuracy of 80.1%, significantly outperforming human experts. Furthermore, the use of saliency maps provides explainable insights, aligning AI decision-making with established clinical biomarkers to foster trust in pediatric clinical settings.

Introduction

In pediatric ophthalmology, the appearance of a swollen optic nerve is a high-stakes diagnostic dilemma. Papilledema—swelling caused by increased intracranial pressure—can signal life-threatening conditions such as brain tumors or meningitis. Conversely, pseudopapilledema is a benign mimic that requires no urgent intervention. Because these conditions appear nearly identical on fundus (back-of-the-eye) images, even trained experts struggle to differentiate them. This study, conducted by researchers at the University of Southern California, introduces an explainable machine learning approach to bridge this diagnostic gap and provide clinicians with a reliable, interpretable decision-support tool.

The Clinical Challenge: Papilledema vs. Pseudopapilledema

The difficulty in diagnosing pediatric papilledema lies in the subtle textural and structural differences between true swelling and anatomical elevation. While deep learning has demonstrated success in general medical imaging, clinical adoption in pediatrics is often hindered by the "black box" nature of neural networks. Doctors are understandably hesitant to rely on algorithms that cannot justify their classifications, especially when the stakes involve invasive procedures or missed emergencies. This research addresses both the accuracy gap and the explainability requirement.

Proposed Methodology

The researchers developed a tri-branch architecture utilizing a DenseNet121 backbone. The pipeline is characterized by three distinct stages designed to mimic and enhance clinical observation:

1. Unsupervised Optic Disc Detection

To focus the model's attention on the relevant anatomy without requiring labor-intensive manual masking, the team implemented an unsupervised localization method. Using intensity thresholding and morphological operations, the system automatically identifies and crops the optic disc from the full fundus image. This preprocessing step ensures the model learns from the most diagnostic region of the eye.

2. Clinically-Inspired Augmentation

Rather than relying on standard random data augmentations, the researchers created three specific "views" for the multi-branch model:

  • Original Crop: The baseline anatomical view.
  • Red-Channel Enhancement: Optimized to highlight hemorrhages and "cotton wool spots," which are primary indicators of nerve damage.
  • Green-Channel Enhancement: Specifically tuned to visualize blood vessel congestion and subtle peripapillary swelling.

3. Multi-View Training and Loss Functions

The model utilizes a dual-loss strategy. Cross-Entropy loss is used for the primary classification task, while Triplet loss is employed to ensure that different augmented views of the same eye are mapped closely in the latent space. This forces the network to learn robust, feature-based representations of the underlying pathology rather than memorizing image noise.

Experimental Setup and Dataset

The study utilized a clinical dataset of 331 pediatric fundus images from 105 subjects (ages 1–17) collected between 2011 and 2021 at Children’s Hospital Los Angeles and the USC Roski Eye Institute. The dataset included:

  • Papilledema (149 images): Confirmed via neuroimaging or lumbar puncture.
  • Pseudopapilledema (182 images): Confirmed via long-term clinical follow-up or normal pressure tests.

To ensure rigorous evaluation, the team employed 10-fold repeated cross-validation. Crucially, images from the same patient were never split between training and testing sets to prevent data leakage and ensure the model's ability to generalize to new patients.

Results and Performance Analysis

The multi-branch model demonstrated superior performance across several metrics:

  • Accuracy: The proposed system achieved 80.1% accuracy, significantly outperforming human experts who averaged 69% on the same dataset.
  • Area Under the Curve (AUC): The model reached an AUC of 0.798, indicating strong discriminative power.
  • Efficiency: By focusing on cropped optic disc images rather than full fundus scans, the model was at least twice as fast to train and execute, making it more suitable for real-time clinical use.
  • Ablation Success: The study confirmed that the multi-branch approach outperformed single-branch models, even when those models had access to the same color-augmented data, proving the value of multi-view representation learning.

Explainability and Clinical Relevance

A core contribution of this work is the use of saliency maps to visualize the AI's decision-making process. By generating "hotspots" on the images, the researchers found that the model’s attention consistently aligned with established clinical biomarkers, such as nasal-side swelling and cotton wool spots. This transparency allows clinicians to verify the AI’s logic, transforming the model from a "black box" into a "human-centered" diagnostic partner.

Conclusion and Future Directions

While the current results are promising, the researchers identified several avenues for future improvement. The dataset, though significant for this rare pediatric condition, remains small for deep learning standards; scaling the database is a primary goal. Additionally, future iterations of the model could incorporate patient metadata—such as age, headache history, or symptoms—to further refine diagnostic accuracy. Finally, incorporating fine-grained manual annotations of specific biomarkers could further enhance the system's explainability, providing even more granular feedback to the treating physician.

Summary updated

Abstract

Papilledema is an ophthalmic neurologic disorder in which increased intracranial pressure leads to swelling of the optic nerves. Undiagnosed papilledema in children may lead to blindness and may be a sign of life-threatening conditions, such as brain tumors. Robust and accurate clinical diagnosis of this syndrome can be facilitated by automated analysis of fundus images using deep learning, especially in the presence of challenges posed by pseudopapilledema that has similar fundus appearance but distinct clinical implications. We present a deep learning-based algorithm for the automatic detection of pediatric papilledema. Our approach is based on optic disc localization and detection of explainable papilledema indicators through data augmentation. Experiments on real-world clinical data demonstrate that our proposed method is effective with a diagnostic accuracy comparable to expert ophthalmologists <sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">1</sup> .
BibTeX
@inproceedings{avramidis2022automating,
  author = {Kleanthis Avramidis and Mohammad Rostami and Melinda Chang and Shrikanth Narayanan},
  title = {{Automating Detection of Papilledema in Pediatric Fundus Images with Explainable Machine Learning}},
  booktitle = {ICIP},
  year = {2022},
  url = {https://openreview.net/forum?id=bgxW2aX1G5},
}
Figure