[Paper Review] Identifying and Categorizing Anomalies in Retinal Imaging Data

Philipp Seeböck, Sebastian M. Waldstein+6 more|arXiv (Cornell University)|Dec 2, 2016

Retinal Imaging and Analysis22 citations

TL;DR

This paper proposes an unsupervised deep learning framework using a Deep Convolutional Autoencoder (DCAE) and One-Class SVM to detect and categorize anomalies in retinal OCT images using only healthy training data. The method achieves a Dice score of 0.55 for anomaly segmentation and enables clinically meaningful clustering of pathologies, with a downstream classification accuracy of 86.6% using learned features.

ABSTRACT

The identification and quantification of markers in medical images is critical for diagnosis, prognosis and management of patients in clinical practice. Supervised- or weakly supervised training enables the detection of findings that are known a priori. It does not scale well, and a priori definition limits the vocabulary of markers to known entities reducing the accuracy of diagnosis and prognosis. Here, we propose the identification of anomalies in large-scale medical imaging data using healthy examples as a reference. We detect and categorize candidates for anomaly findings untypical for the observed data. A deep convolutional autoencoder is trained on healthy retinal images. The learned model generates a new feature representation, and the distribution of healthy retinal patches is estimated by a one-class support vector machine. Results demonstrate that we can identify pathologic regions in images without using expert annotations. A subsequent clustering categorizes findings into clinically meaningful classes. In addition the learned features outperform standard embedding approaches in a classification task.

Motivation & Objective

• To detect diagnostically relevant anomalies in retinal OCT images without relying on annotated pathological data.
• To develop a method that identifies novel biomarkers by learning from only healthy retinal images.
• To categorize detected anomalous regions into clinically meaningful subtypes using clustering.
• To evaluate the discriminative power of the learned features in a downstream classification task.

Proposed method

• A multi-scale Deep Convolutional Autoencoder (DCAE) is trained on healthy OCT patches to learn hierarchical feature representations.
• Features from two DCAE scales (32×32 and downsampled 128×32) are concatenated and fed into a Denoising Autoencoder to produce a 256D feature vector per superpixel.
• A One-Class SVM is trained on the 256D features of healthy superpixels to define a decision boundary for normal vs. anomalous regions.
• Anomalous superpixels are further segmented into subtypes using spherical K-means clustering with cosine distance and the Davies-Bouldin index for optimal cluster count.
• The framework enables pixel-wise anomaly segmentation and subsequent categorization into biologically interpretable clusters.
• Performance is evaluated via Dice score, classification accuracy on IRC/SRF/other, and expert validation of cluster interpretations.

Experimental results

Research questions

• RQ1 Can an unsupervised model detect clinically relevant anomalies in retinal OCT images using only healthy training data?
• RQ2 Can the detected anomalous regions be meaningfully categorized into distinct pathological subtypes through clustering?
• RQ3 Do the learned features from the DCAE have sufficient discriminative power for downstream classification tasks?
• RQ4 How do the proposed methods compare to conventional PCA-based feature extraction in anomaly detection and classification?

Key findings

• The proposed DCAEent model achieved a Dice score of 0.55 for anomaly segmentation, significantly outperforming PCA-based baselines (0.33 and 0.32).
• The One-Class SVM with DCAE features successfully segmented anomalous regions with high precision (0.53) and recall (0.58).
• Spherical K-means clustering with 7 clusters yielded the lowest Davies-Bouldin index and enabled clinically interpretable categorization of anomalies.
• Experts identified cluster 1 as RPE with bright horizontal edges, cluster 4 as intraretinal fluid, and clusters 3 and 5 as subretinal fluid, confirming clinical relevance.
• The DCAEent features achieved a mean classification accuracy of 86.6% in a 3-class task (IRC, SRF, other), outperforming PCA-based models (73.4% and 73.3%).
• t-SNE visualization confirmed that the learned feature space captures meaningful semantic structure, with clusters corresponding to distinct anatomical and pathological patterns.