Multimodal Symbolic Regex Description Patterns for Multiclass Macular Lesion Classification

Accurate classification of macular lesions in retinal fundus images supports the early detection and monitoring of diabetic retinopathy and age related macular degeneration. Deep learning methods have become well established in these tasks, although they tend to be computationally costly and difficult to interpret. In this paper a classical machine learning pipeline for the classification of macular lesion patches is proposed, requiring no deep neural network component. The clinical target comprises three edema related lesion classes (hard exudates, micro hemorrhages, and cotton wool exudates), one age related macular degeneration indicator (drusen), and a healthy control class. The dataset contains 2384 RGB patches of size 32×32 pixels annotated by an ophthalmology specialist, released with patch to image mappings but without patient level metadata. The proposed method, a Multimodal Symbolic Retinal Descriptor with Grammar Induced Regular Expressions and an XGBoost classifier (MSRD-XGB), integrates a compact handcrafted descriptor with a symbolic string encoding stage that converts each patch into quantized and binary symbolic modalities sampled through complementary spatial paths. A finite regex grammar then induces candidate structural features, which are scored on training subsets, stabilized by stratified bootstrap, pruned by correlation, and selected jointly with the final XGBoost configuration on the validation split. On the validation selected split the model reached 0.8562 multiclass Matthews correlation coefficient (MCC) on test, with 0.8973 accuracy and 0.7926 macro F1. The mean test MCC was 0.8592 ± 0.0143 across 30 repeated patch level clean splits and 0.7961 ± 0.0190 across 30 stricter source image splits. An external transfer protocol derived from the Indian Diabetic Retinopathy Image Dataset segmentation masks yielded 0.8300 MCC. These results suggest that the proposed pipeline is a useful and computationally efficient option for low resolution retinal patch classification.

Authors: