A Radiomic-Based Machine Learning System to Diagnose Age-Related Macular Degeneration from Ultra-Widefield Fundus Retinography

Abstract

The present study was conducted to investigate the potential of radiomics to develop an explainable AI-based system to be applied to ultra-widefield fundus retinographies (UWF-FRTs) with the objective of predicting the presence of the early signs of Age-related Macular Degeneration (AMD) and stratifying subjects with low- versus high-risk of AMD. The ultimate aim was to provide clinicians with an automatic classifier and a signature of objective quantitative image biomarkers of AMD. The use of Machine Learning (ML) and radiomics was based on intensity and texture analysis in the macular region, detected by a Deep Learning (DL)-based macular detector. Two-hundred and twenty six UWF-FRTs were retrospectively collected from two centres and manually annotated to train and test the algorithms. Notably, the combination of the ML-based radiomics model and the DL-based macular detector reported 93% sensitivity and 74% specificity when applied to the data of the centre used for external testing, capturing explainable features associated with drusen or pigmentary abnormalities. In comparison to the human operator's annotations, the system yielded a 0.79 Cohen κ, demonstrating substantial concordance. To our knowledge, these results are the first provided by a radiomic approach for AMD supporting the suitability of an explainable feature extraction method combined with ML for UWF-FRT.

Keywords: age-related macular degeneration (AMD); artificial intelligence (AI); deep learning (DL); detection; explainability; fundus retinography (FRT); image segmentation; machine learning (ML); radiomics; ultra-widefield (UWF).

Figure 1

DSCs for the comparison between DL-based macular detector’s predictions and human operator’s manual ROIs, on the training, validation and internal testing sets.

Figure 2

DSCs for the comparison between DL-based macular detector’s predictions and human operator’s manual ROIs, on the external testing set.

Figure 3

DSCs for the comparison between DL-based macular detector’s predictions and human operator’s manual ROIs, on the external testing, split according to whether the detector failed (above, Automatic Mean) or succeeded (below, Model Prediction) in detecting the macula.

Figure 4

Predictions of the DL-based macular detector, on representative UWF-FRT from the internal testing set. The predictions (in red) are compared with the reference standard human operator’s manual ROIs (in blue).

Figure 5

Predictions of the DL-based macular detector, on representative UWF-FRT from the external testing set. The predictions (in red) are compared with the reference standard human operator’s manual ROIs (in blue).

Figure 6

(A) ROC Curve for each of the three ensembles (represented in blue, red and yellow) of random forest classifiers (from Internal Testing). (B) ROC Curve for each of the three ensembles (represented in blue, red and yellow) of support vector machine classifiers (from Internal Testing).

Figure 7

Violin and box plots for the top 10 radiomic predictors. Violin and box plots of “AMD” and “Negative” classes are in red and green, respectively.

Figure 8

Representative examples from external testing, one example for each classification outcome. The ML-based radiomics model’s predictions are obtained from ROIs obtained by the DL-based macular detector (in red). For each ROI it is also reported the reference standard human operator’s manual ROIs (in blue). (A) Example of True Positive (“AMD” correctly classified as “AMD”). (B) Example of True Negative (“Negative” correctly classified as “Negative”). (C) Example of False Positive (“Negative” incorrectly classified as “AMD”). (D) Example of False Negative (“AMD” incorrectly classified as “Negative”).