Search

Introduction: In this retrospective cohort study, we wanted to evaluate the performance and analyze the insights of an artificial intelligence (AI) algorithm in detecting retinal fluid in spectral-domain OCT volume scans from a large cohort of patients with neovascular age-related macular degeneration (AMD) and diabetic macular edema (DME). Methods: A total of 3’981 OCT volumes from 374 patients with AMD and 11’501 OCT volumes from 811 patients with DME, acquired with Heidelberg Spectralis OCT device (Heidelberg Engineering Inc., Heidelberg, Germany) between 2013 and 2021. Each OCT volume was annotated for the presence or absence of intraretinal fluid (IRF) and subretinal fluid (SRF) by masked reading center graders (ground truth). The performance of an already published AI-algorithm to detect IRF, SRF separately and a combined fluid detector (IRF and/or SRF) of the same OCT volumes was evaluated. An analysis of the sources of disagreement between annotation and prediction and their relationship to central retinal thickness was performed. We computed the mean areas under the curves (AUC) and under the precision-recall curves (AP), accuracy, sensitivity, specificity and precision. Results: The AUC for IRF was 0.92 and 0.98, for SRF 0.98 and 0.99, in the AMD and DME cohort, respectively. The AP for IRF was 0.89 and 1.00, for SRF 0.97 and 0.93, in the AMD and DME cohort, respectively. The accuracy, specificity and sensitivity for IRF was 0.87, 0.88, 0.84, and 0.93, 0.95, 0.93, and for SRF 0.93, 0.93, 0.93, and 0.95, 0.95, 0.95 in the AMD and DME cohort respectively. For detecting any fluid, the AUC was 0.95 and 0.98, the accuracy, specificity and sensitivity was 0.89, 0.93, 0.90 and 0.95, 0.88 and 0.93, in the AMD and DME cohort, respectively. False positives were present when retinal shadow artifacts and strong retinal deformation were present. False negatives were due to small hyporeflective areas in combination with poor image quality. The combined detector correctly predicted more OCT volumes than the single detectors for IRF and SRF, 89.0% versus 81.6% in the AMD and 93.1% versus 88.6% in the DME cohort. Discussion/Conclusion: The AI-based fluid detector achieves high performance for retinal fluid detection in a very large dataset dedicated to AMD and DME. Combining single detectors provides better fluid detection accuracy than considering the single detectors separately. The observed independence of the single detectors ensures that the detectors learned features particular to IRF and SRF.

Purpose: To assess the potential of machine learning to predict low and high treatment demand in real life in patients with nAMD, RVO and DME treated according to a TER. Design: Retrospective cohort study. Participants: 377 eyes (340 patients) with nAMD, 333 eyes (285 patients) with RVO or DME, treated with anti-VEGF according to a predefined TER protocol during 2014-2018. Methods: Eyes were grouped by disease into low, moderate and high treatment demanders, defined by the average treatment interval (low: ≥10 weeks, high: ≤5 weeks, moderate: remaining eyes). Two Random Forest models were trained to predict the probability of the long-term treatment demand of a new patient. Both models use morphological features automatically extracted from the OCT volumes at baseline and after two consecutive visits, as well as patient demographic information. Evaluation of the models included a 10 cross-validation ensuring that no patient was present in both training (nAMD: ~339, RVO & DME: ~300) and test sets (nAMD: ~38, RVO & DME: ~33). Main Outcome Measures: Mean Area under the Curve (AuC) of both models, contribution to the prediction and statistical significance of the input features. Results: Based on the first three visits, it was possible to predict a low and a high treatment demander in nAMD and RVO & DME eyes with similar accuracy. Prediction performance for low and high treatment demand showed similar performance levels across nAMD and RVO & DME eyes. For nAMD, 127 low, 42 high- and 208 moderate demanders were identified, and for RVO & DME 61 low-, 50 high- and 222 moderate demanders. The nAMD trained models yielded mean AuCs of 0.79 and 0.79 over the 10 folds for low and high demanders, respectively. Models on RVO & DME showed similar results with a mean AuC of 0.76 and 0.78 for low and high demanders, respectively. Even more importantly, this study reveals that it is possible to predict reasonably well low demands at the first visit, before the first injection. Conclusions: Machine learning classifiers can predict treatment demand and may assist in establishing patient specific treatment plans in the near future.

Retinoblastoma and uveal melanoma are fast spreading eye tumors usually diagnosed by using 2D Fundus Image Photography (Fundus) and 2D Ultrasound (US). Diagnosis and treatment planning of such diseases often require additional complementary imaging to confirm the tumor extend via 3D Magnetic Resonance Imaging (MRI). In this context, having automatic segmentations to estimate the size and the distribution of the pathological tissue would be advantageous towards tumor characterization. Until now, the alternative has been the manual delineation of eye structures, a rather time consuming and error-prone task, to be conducted in multiple MRI sequences simultaneously. This situation, and the lack of tools for accurate eye MRI analysis, reduces the interest in MRI beyond the qualitative evaluation of the optic nerve invasion and the confirmation of recurrent malignancies below calcified tumors. In this manuscript, we propose a new framework for the automatic segmentation of eye structures and ocular tumors in multi-sequence MRI. Our key contribution is the introduction of a pathological eye model from which Eye Patient-Specific Features (EPSF) can be computed. These features combine intensity and shape information of pathological tissue while embedded in healthy structures of the eye. We assess our work on a dataset of pathological patient eyes by computing the Dice Similarity Coefficient (DSC) of the sclera, the cornea, the vitreous humor, the lens and the tumor. In addition, we quantitatively show the superior performance of our pathological eye model as compared to the segmentation obtained by using a healthy model (over 4% DSC) and demonstrate the relevance of our EPSF, which improve the final segmentation regardless of the classifier employed.