Explainable machine learning framework for cataracts recognition using visual features

Abstract

Cataract is the leading ocular disease of blindness and visual impairment globally. Deep neural networks (DNNs) have achieved promising cataracts recognition performance based on anterior segment optical coherence tomography (AS-OCT) images; however, they have poor explanations, limiting their clinical applications. In contrast, visual features extracted from original AS-OCT images and their transform forms (e.g., AS-OCT-based histograms) have good explanations but have not been fully exploited. Motivated by these observations, an explainable machine learning framework to recognize cataracts severity levels automatically using AS-OCT images was proposed, consisting of three stages: visual feature extraction, feature importance explanation and selection, and recognition. First, the intensity histogram and intensity-based statistical methods are applied to extract visual features from original AS-OCT images and AS-OCT-based histograms. Subsequently, the SHapley Additive exPlanations and Pearson correlation coefficient methods are applied to analyze the feature importance and select significant visual features. Finally, an ensemble multi-class ridge regression method is applied to recognize the cataracts severity levels based on the selected visual features. Experiments on a clinical AS-OCT-NC dataset demonstrate that the proposed framework not only achieves competitive performance through comparisons with DNNs, but also has a good explanation ability, meeting the requirements of clinical diagnostic practice.

Keywords: Anterior segment optical coherence tomography; Explainable; Machine learning; Nuclear cataract; Visual feature.

Fig. 1

Three representative AS-OCT images and their corresponding AS-OCT-based histograms for three NC severity levels: a normal, b mild, and c severe. The AS-OCT-based histograms are built by counting the pixel values in the AS-OCT images. The pixel numbers of each interval are also referred to as intensity. In comparison, the original AS-OCT images of different NC severity levels are very similar, but their AS-OCT based histograms are significantly different

Fig. 2

Flowchart of the proposed explainable machine learning framework. Given an AS-OCT image, a deep segmentation network was first applied to segment the nucleus region from AS-OCT images automatically. Secondly, 23 histogram-based statistical features from the AS-OCT-based histogram and four clinical intensity-based statistical features from the original AS-OCT image were extracted. Subsequently, the relative importance of the features and select an informative feature set based on SHAP and PCC was analyzed. Finally, the EMRR to recognize the cataracts severity level was proposed

Fig. 3

RR models for different NC severity levels: a normal, b mild, and c severe. The EMRR first calculates the $P(y_i)$ for each level based on the corresponding model output $y_i$. Subsequently, it selects the level with the largest $P(y_i)=0.525$ as the final output, which is severe

Fig. 4

SHAP values between the 27 visual features and NC severity levels: a normal, b mild, c severe, and d overall. The visual features contributed differently to recognizing different NC severity levels, and the height of each denotes their importance

Fig. 5

PCC matrix between features. The coefficients ranged from -1 to 1, indicating the degree of correlation. Values close to -1 and 1 indicate a high correlation, whereas values close to 0 indicate a low correlation

Fig. 6

Training loss and ACC of different DNNs with a learning rate of 0.001

Fig. 7

Confusion matrices of the EMRR, LR, and four other DNNs on AS-OCT-NC dataset with a learning rate of 0.001

Fig. 8

Performance comparison of machine learning methods based on histogram-based statistical features in different intensity ranges

Fig. 9

Performance comparison of machine learning methods based on histogram-based statistical features in different intensity intervals