Interpretable Deep Learning for OCT-Based Diagnosis of Vitreoretinal Lymphoma Versus Uveitis

Abstract

Purpose

To develop and validate an interpretable deep learning model that classifies optical coherence tomography (OCT) scans as vitreoretinal lymphoma (VRL) or non-infectious uveitis (NIU)—intermediate, posterior, or panuveitis—and visualizes differentiating pathological OCT features.

Methods

This cross-sectional study included 45 patients with VRL and 52 with NIU who underwent SPECTRALIS OCT imaging. OCT B-scans were analyzed with a custom-designed deep convolutional neural network classification model combined with gradient-weighted class activation mapping (Grad-CAM). Interpretability was assessed using Grad-CAM heat maps that visualize disease-specific OCT features. We evaluated diagnostic performance at the eye level using the area under the receiver operating characteristic curve (AUROC), area under the precision–recall curve (AUPRC), accuracy, F1 score, precision, sensitivity, specificity, and kappa.

Results

The model differentiated VRL from NIU at the eye level, with mean ± SD scores of 76.0 ± 8.0 for AUROC, 79.4 ± 10.4 for AUPRC, 75.4 ± 13.3 for accuracy, 74.9 ± 13.7 for F1 score, 73.9 ± 19.3 for precision, 77.7 ± 13.6 for sensitivity, 73.4 ± 20.7 for specificity, and 51.0 ± 26.6 for kappa. Grad-CAM features of VRL included preretinal deposits and changes in the retinal pigment epithelium. In contrast, NIU features included denser vitreous cells and opacities, epiretinal membrane, intraretinal cells, and choroidal and scleral abnormalities consistent with NIU pathophysiology.

Conclusions

A deep learning model with interpretability can reliably differentiate VRL from NIU, highlighting disease-relevant OCT features.

Translational Relevance

The proposed deep learning method may serve as a non-invasive triage tool by identifying patients who need confirmatory pars plana vitrectomy.

Introduction

Vitreoretinal lymphoma (VRL) is a rare but aggressive intraocular malignancy and a subset of primary central nervous system lymphoma, most often (95%) categorized as diffuse large B-cell lymphomas^1–3 It typically involves the retina and vitreous and mimics chronic uveitis, frequently leading to delays in diagnosis.^4,5 Most studies show that the median age of individuals diagnosed with VRL is 60 years, and the median survival time from diagnosis is 33 months.^1,3,5,6 Delayed diagnosis of VRL is associated with high mortality, largely due to progressive central nervous system involvement.^3,5,7,8 Hence, there is a critical need for timely and accurate diagnosis. Importantly, to diagnose VRL, a liquid biopsy (aqueous, vitreous, or cerebrospinal fluid) or solid biopsy (chorioretinal or brain) must be performed for cytologic examination.^9–11 This is most commonly performed through a pars plana vitrectomy with vitreous biopsy. Once diagnosed, VRL is typically treated with a combination of systemic chemotherapy and local ocular therapy, such as intravitreal chemotherapy or radiotherapy.⁵

Several optical coherence tomography (OCT) features have been reported in VRL, including intraretinal infiltrates, subretinal and sub-retinal pigment epithelium (sub-RPE) infiltrates, vitreous cells, pre-retinal deposits, confluent RPE detachments, and the absence of an epiretinal membrane (ERM).^11–13 However, overlap with the clinical features of non-infectious uveitis (NIU)—intermediate, posterior, or panuveitis—remains a major challenge.^1,11 Previously, OCT-based diagnosis of VRL relied on manual grading and rule-based approaches using differing combinations of OCT features.¹¹ However, traditional diagnostic approaches based on expert-defined OCT features remain prone to intergrader variability, indicating a gap that deep learning methods can address. The aim of our study was to develop a deep convolutional neural network (CNN) for the differentiation of VRL and NIU using OCT B-scans, with interpretability provided through gradient-weighted class activation mapping (Grad-CAM)¹⁴ heatmaps.

Methods

Data Acquisition

The data in support of this study were collected from the Casey Eye Institute, Oregon Health & Science University (OHSU), Portland, OR. This study was approved by the OHSU institutional review board (IRB) and was exempt from informed consent given its retrospective nature. Biopsy-confirmed VRL and active NIU patients ≥ 45 years of age with good-quality OCT B-scans at the initial visit identified through the International Classification of Diseases, Tenth Revision, codes were included in this study. Good-quality OCT B-scans are well centered, with complete retinal structure and sufficient signal quality for clinical analysis. Exclusion criteria included infectious uveitis; history of eye surgery within the last 12 months; active use of systemic therapy, including oral corticosteroids, for disease management; and poor-quality OCT B-scans. Local corticosteroids were permitted, as the majority of VRL patients are referred to tertiary care centers on at least topical therapy. Poor-quality images typically exhibit reduced signal strength, increased speckle noise, and insufficient retinal structure information, making clinical interpretation difficult. For each patient, one or two eyes were scanned using a SPECTRALIS OCT device (Heidelberg Engineering, Heidelberg, Germany). The OCT scan pattern uses fast and posterior pole volumetric raster scan protocols with fields of view of 20° and 30°. For the 20° raster scan protocol, each B-scan had a resolution of 1024 × 496 pixels; for the 30° raster scan protocol, each B-scan had a resolution of 1536 × 496 pixels. Volumetric acquisitions were performed using 19-line, 25-line, or 61-line raster scans.

Data Inputs

In clinical practice, volumetric OCT scans often vary in size, which complicates model training. To address this issue, we used B-scans extracted from the OCT volume, providing a large number of samples for training. Another problem is variability in B-scan sampling density across different scan protocols. To address this problem, OCT B-scans acquired with different scan protocols were resized by downsampling to a uniform resolution of 1024 × 496 pixels before feeding them into the model. The fixed size allows for uniform batching during training. Although this resizing involves some downsampling, it strikes a balance between preserving B-scan quality and facilitating model training. To prevent data leakage, all B-scans from the same eye were assigned exclusively to a single fold; therefore, no B-scans from the same eye appeared in any of the training, validation, or testing folds.

Classification Network Using Deep Learning

The architecture of the proposed deep learning model for classifying VRL versus NIU is illustrated in Figure 1. The model uses a deep CNN comprised of stacked residual squeeze-and-excitation (SE) blocks (Fig. 1C), followed by fully connected layers for classification. To enhance feature extraction, we incorporated residual connections inspired by the ResNet architecture and integrated channel-wise attention using the SE mechanism,¹⁵ as shown in Figures 1A and 1B, respectively. Residual connections are used to mitigate the vanishing gradient problem and enable deeper feature learning by allowing the network to learn identity mappings alongside nonlinear transformations (Fig. 1A). The SE block (Fig. 1B) function adaptively reweights each feature channel (based on learned importance) to emphasize more informative features and suppress less informative ones, thereby improving the ability of the model to focus on disease-relevant signals.

Figure 1.

Architecture of the CNN used for classification in this study. (A) Residual block used to learn residual feature representations. (B) SE block that performs channel-wise feature recalibration. (C) Residual SE block combining residual connections with SE-based attention. (D) Complete 2D CNN architecture consisting of stacked residual SE blocks, max pooling layers, and a classification head. The classification head includes adaptive average pooling, flattening, a fully connected layer (FC1), dropout, a final fully connected layer (FC2), and a SoftMax activation for multi-class prediction.

In an input feature map of shape [B, C, H, W], B is the batch size, C is the number of channels, and height (H) and width (W) are spatial dimensions. The SE block first applies global average pooling to squeeze each channel into a single value (channel descriptor), resulting in the shape of [B, C, 1, 1]. This vector is flattened to [B, C], passed through two fully connected layers: the first fully connected layer (FC1) reduces it to [B, C/r] (with a reduction ratio of r = 16), followed by rectified linear unit (ReLU) activation; the second fully connected layer (FC2) expands it back to [B, C] and applies a sigmoid activation to produce channel-wise weights in the range (0, 1). These weights are reshaped to [B, C, 1, 1] and multiplied element-wise with the original feature map (Fig. 1B, modulation line) to recalibrate the importance of each channel, producing a refined output of shape [B, C, H, W].

By combining residual learning and SE-based channel attention, we constructed a residual SE block (Fig. 1C) based on the SE-ResNet module proposed in the original squeeze-and-excitation networks paper.¹⁵ In their study, the authors demonstrated improved performance when integrating SE blocks into ResNet-50, with only a slight increase in computational cost. We adopted this residual SE block structure in our model, stacking multiple residual SE blocks with two-dimensional (2D) max pooling layers in between, followed by a classification head consisting of adaptive average pooling, flattening, fully connected layers (FC1, FC2), and SoftMax¹⁶ activation (Fig. 1D).

To provide clinically meaningful interpretability and support clinical decision making, we utilized Grad-CAM. This popular interpretability technique highlights (via heatmap visualization) the regions of the OCT B-scan contributing most to the prediction of the model, thereby providing a clearer understanding and promoting transparency in decision making. We used the feature maps from the final convolutional block to generate class-specific activation maps, enabling the localization of pathologic features that contribute to the model decision.

Classification Performance Evaluation

We performed eye-level classification by aggregating B-scan level predictions. To achieve this, we assigned each eye to VRL (positive class) or NIU (negative class) based on the average SoftMax probability (ASP) vectors across all B-scans within the OCT volume. The class with the highest ASP was then assigned the final eye-level label. To assess the diagnostic accuracy of our model in differentiating VRL from NIU, we employed a comprehensive set of performance metrics that included area under the receiver operating characteristic curve (AUROC), which showed that a model with a higher AUROC indicates that it reliably assigns higher scores to true VRL cases than to NIU cases. Area under the precision–recall curve (AUPRC) confirmed that the model maintains high detection rates without excessive false positives. Accuracy showed the overall proportion of correctly classified cases (overestimating performance if one class dominates). The F1 score balanced precision and recall when working with an imbalanced dataset. Precision was used to calculate the proportion of predicted positive cases (true VRL), which is essential for minimizing unnecessary invasive biopsies (e.g., pars plana vitrectomy) in patients without VRL. Sensitivity was used to calculate the proportion of actual VRL cases correctly identified. This metric is essential for patients predicted to have VRL to receive timely intervention. Specificity was used to calculate the proportion of NIU cases correctly identified, which helps to avoid inappropriate treatment. Cohen's kappa was used to measure the level of agreement between two raters (model predictions and the ground truth labels) while accounting for the possibility of agreement occurring by chance. The evaluation metrics, including accuracy, F1 score, precision, sensitivity, and specificity, were calculated as defined in Equation 1.

$$\begin{array}{ccc}\hfill Accuracy& =& \frac{TP+TN}{TP+TN+FP+FN}\hfill \\ \hfill F1& =& \frac{2\times TP}{2\times TP+FP+FN},\hfill \\ \hfill Precision& =& \frac{TP}{TP+FP},\hfill \\ \hfill \mathit{Sensitivity}& =& \frac{TP}{TP+FN},\hfill \\ \hfill \mathit{Specificity}& =& \frac{TN}{TN+FP}\hfill \end{array}$$ (1)

where TP (true positives) is the number of correctly identified positives, TN (true negatives) is the number of correctly identified negatives, FP (false positives) represents incorrectly identified positives, and FN (false negatives) represents incorrectly identified negatives.

We used five-fold stratified cross-validation to ensure robust and unbiased evaluation. In each fold, 20% of the cases were reserved for testing, and the remaining 80% were split into 68% for training and 12% for validation. Stratification preserved class balance across all subsets, and no case appeared in more than one split per fold, avoiding data leakage.

Implementation Details

The models were implemented in Python using the PyTorch deep learning framework (GPU-enabled system). To ensure reproducibility, we fixed random seeds during training. A grid search strategy was used to identify optimal hyperparameters, focusing on learning rate, batch size, and dropout. The models were trained with the cross-entropy¹⁷ loss function and optimized using the Adam algorithm.¹⁸ A learning rate scheduler was used that decays the learning rate by a gamma factor after a defined step size, promoting stable convergence. We employed early stopping based on validation accuracy, saving the checkpoint with the best validation performance (highest accuracy) for each model. To reduce overfitting and improve generalization, we applied data augmentation such as random horizontal flipping and random rotations. The folds model was trained on a high-performance computing system equipped with two GeForce GTX 1070 Ti GPUs (NVIDIA, Santa Clara, CA), enabling fast training and inference.

Results

We obtained data from 111 patients, including 52 with VRL and 59 with NIU. Of the 52 VRL patients, 45 eyes were of sufficient quality for inclusion, and 52 eyes from the 59 NIU patients met the quality criteria. The mean ages of the VRL and NIU patients were 70.4 years and 57.8 years, respectively (P < 0.001). A total of 20 VRL patients (47.6%) and 29 NIU patients (58.0%) were female (P = 0.4). NIU cases included those with confirmed diagnoses of birdshot chorioretinopathy (15), ocular sarcoidosis (14), Vogt–Koyanagi–Harada disease (14), pars planitis (9), sympathetic ophthalmia (3), multifocal choroiditis (2), and idiopathic panuveitis (2).

Quantitative Evaluation

To quantitatively evaluate the diagnostic performance of the proposed classification model, we determined the mean values of evaluation metrics across five test folds (Table). On average, the model demonstrated reliable performance, with all key metrics in the table indicating strong discriminative ability between the two classes. Out of all of the folds, fold 3 achieved the highest AUROC (0.89), demonstrating strong separation between VRL and NIU (Fig. 2A). Similarly, we achieved an AUPRC of 0.93 in the precision–recall curve (Fig. 2B), showing a good balance between precision and recall in identifying VRL cases. Together, these metrics show that our model could distinguish eyes with VRL from those with NIU. We summarized the eye-level predicted ASP of VRL, grouped by true diagnosis. True VRL eyes showed consistently higher ASP than true NIU eyes, indicating that the model achieved sufficient eye-level separation (Fig. 3).

Table.

Results of Automated Differentiation of VRL From Non-Infectious Intermediate, Posterior, or Panuveitis Obtained Through 5-Fold Stratified Cross-Validation

Metric	VRL Diagnosis (%), Mean ± SD (95% CI)
Accuracy	75.4 ± 13.3 (58.9–92.0)
F1 score	74.9 ± 13.7 (57.8–92.0)
Precision	73.9 ± 19.3 (49.8–97.9)
Sensitivity	77.7 ± 13.6 (60.8–94.6)
Specificity	73.4 ± 20.7 (47.7–99.2)
AUROC	76.0 ± 8.0 (65.0–86.0)
AUPRC	79.4 ± 10.4 (66.0–92.0)
kappa	51.0 ± 26.6 (18.0–84.0)

Figure 2.

Eye-level receiver operating characteristic (ROC) and precision–recall (PR) curves for distinguishing VRL from NIU (intermediate, posterior, or panuveitis).  (A) The ROC curve demonstrates robust classification performance with a high AUROC of 0.89 for fold 3, indicating good separation between the two conditions. (B) The PR curve demonstrates a good balance between precision and recall, with an AUPRC of 0.93 for fold 3. Together, these results confirm that the proposed model has learned disease-relevant features and provides reliable eye-level classification of VRL versus NIU. CI, confidence interval.

Figure 3.

Eye-level distribution of ASP for VRL and NIU (intermediate, posterior, or panuveitis). The box plot shows the ASP for VRL in true VRL eyes and in true NIU eyes. Most VRL eyes show high predicted VRL ASP. In contrast, NIU eyes show low VRL ASP, indicating sufficient eye-level separation with a small number of uncertain cases.

Representative Cases

To validate the predictions of our model and interpret its decision-making potential, we applied Grad-CAM to the output feature maps of the final convolutional block to generate class-specific activation maps. The model provides visual explanations of its predictions through Grad-CAM heatmaps. These heatmaps highlight the regions of each OCT B-scan that contribute most to the output of the model. The pathological OCT features highlighted by the model are evident in the VRL (Fig. 4) and NIU (Fig. 5) cases. By localizing disease-relevant features, this approach enhances interpretability and supports clinical understanding of the decision-making process of the model.

Figure 4.

A representative case of a 72-year-old female diagnosed with primary VRL, showing OCT features of the left eye. (A) Input OCT B-scan shows preretinal deposits (blue arrow) and RPE changes (yellow arrow). (B) The corresponding Grad-CAM heatmap overlays on the input OCT B-scan highlight these regions as the most predictive of the model decision. (C) A second input OCT B-scan shows similar findings of preretinal deposits (blue arrow) and RPE changes (yellow arrow). (D) The corresponding Grad-CAM heatmap overlays on the input OCT B-scan again highlight these features, confirming their contribution to the prediction of the model.

Figure 5.

Representative case of a 60-year-old female diagnosed with birdshot chorioretinopathy showing OCT features of the left eye. (A) Input OCT B-scan shows vitreous cells/opacities (green arrow), epiretinal membrane (red arrow), and choroidal and scleral abnormalities (predominantly thickening, orange arrow). (B) The corresponding Grad-CAM heatmap overlays on the input OCT B-scan highlight these regions as the most predictive of the model decision. (C) A second input OCT B-scan shows vitreous cells/opacities (green arrow) and choroidal and scleral abnormalities (predominantly thickening, orange arrow). (D) The corresponding Grad-CAM heatmap overlays on the input OCT B-scan again highlight these features, confirming their contribution to the prediction of the model.

Discussion

In this study, we proposed and validated an automated diagnostic framework (based on a deep CNN) for differentiating VRL from NIU using OCT B-scans. The model utilizes residual SE blocks with 2D max pooling layers in between, followed by a classification head. This new combination in our task enables the model to learn deeper features and focuses on the most informative features (suppressing less relevant ones). On average, the model demonstrated reliable diagnostic performance across five-fold stratified cross-validation with Grad-CAM heatmaps that highlighted disease-relevant features (Figs. 4, 5). These findings suggest that the model learns clinically meaningful OCT features, supporting its potential utility in distinguishing VRL from NIU.

Previous studies have attempted to differentiate VRL from NIU using OCT-based features. Guan et al.¹¹ used a rule-based approach and reported preretinal deposits as a highly specific differentiating feature of VRL, and RPE changes were identified as a sensitive feature. El Zein et al.¹³ manually reviewed and compared VRL and NIU OCT features, concluding that the absence of an ERM was a distinguishing feature of VRL. Artificial intelligence has also been used to aid in the diagnosis of VRL. A machine learning model by Gozzi et al.² used an XGBoost-based classification model to differentiate VRL from anterior uveitis using anterior segment OCT (AS-OCT) images. This study focused on traditional machine learning applied only to AS-OCT, an imaging modality not commonly used in retina or uveitis practices at this time. Lin et al.¹⁹ used a few-shot deep learning approach and reported only VRL features on the OCT B-scan. Our study corroborated that preretinal deposits and RPE changes were sensitive in the detection of VRL, and the presence of an ERM was found to be a distinguishing feature of NIU (Figs. 4, 5). We also found that choroidal/scleral abnormalities were more common in NIU, which is expected given the known phenomenon of sclerochoroidal thickening with active posterior and panuveitis. It is also important to note that younger patients are known to have increased choroidal thickness compared to those who are older. Given that patients with VRL were older than those with NIU in our study, this may have contributed to our findings. We also performed eye-level classification and displayed B-scan features using Grad-CAM heatmaps. This would allow clinicians to visualize the features most predictive of the model decision. Our approach is transparent and clinically applicable for further investigation of suspected VRL patients and would benefit from validation using larger datasets.

This study has several limitations. First, due to the rarity of the disease, our dataset was relatively small. This precluded the use of volumetric OCT scans for training a three-dimensional (3D) CNN. A further challenge was the variability in the number of B-scans across the volume, which limited the feasibility of consistent 3D model training and testing. Second, we used retrospective data from a single institution and a single OCT device, which may limit the generalizability of our findings. To address this, we are currently expanding the study to an IRB-approved multicenter and multiple-device cohort. Such external validation will be required before clinical deployment to confirm the generalizability and clinical reliability of the proposed method, which will enhance the translational relevance of this study. Third, our NIU cohort had limited uveitic diagnoses. We excluded infectious etiologies such as syphilis and tuberculosis given their diverse presentations and the presence of serologic testing to evaluate for these entities, but we hope to incorporate them in future studies. VRL can also resemble placoid uveitides and other white dot syndromes, such as multiple evanescent white dot syndrome, acute zonal occult outer retinopathy, and punctate inner choroidopathy, which we hope to incorporate into new models. The inclusion of intermediate uveitis cases was to evaluate if the model could detect differences in reflectivity in the posterior vitreous portion of the scans, but differences in the cellularity of each B-scan may have impacted the interpretation and accuracy of the model. Finally, our analysis was limited to structural OCT data. Our approach could complement existing methods as part of a multimodal diagnostic pathway, serving as an initial non-invasive tool that examines retinal features in OCT B-scans. This approach reflects real-world clinical practice, where combining multiple sources of evidence can guide clinical decision making regarding whether a diagnostic procedure should be pursued in a patient.

In future work, we also plan to include OCT angiography (OCTA), which may enhance diagnostic accuracy by providing additional information on retinal vasculature. OCTA provides detailed information on retinal vasculature, which can aid in diagnosis by visualizing the vasculature and its structural features. VRL and NIU groups were not age matched, limiting our ability to perform age-adjusted or sensitivity analyses. This may have implications for sclerochoroidal thickness, as it is well known that the choroid thins with older age. Despite this, the current model was able to highlight disease-relevant features, consistent with previous clinical OCT findings. Future studies with larger, age-matched cohorts will allow for sensitivity analyses to better separate disease-relevant OCT features from age-related structural changes. Furthermore, OCT image quality remains a practical limitation. To automate this process, we are developing an automated tool that will flag low-quality scans and will ultimately support clinical deployment by ensuring that the diagnostic model analyzes only reliable OCT scans.

Conclusions

We have proposed a deep learning model to differentiate VRL from NIU using OCT data as input. Our model achieved reliable diagnostic performance and highlighted key features through Grad-CAM heatmaps, which corresponded with previously reported clinical OCT findings. Our proof-of-concept model highlights the potential of OCT-based diagnosis as a non-invasive tool to identify patients suspected of VRL before a confirmatory biopsy.