<![CDATA[Early detection of Microaneurysms (MAs) is vital for diagnosing Diabetic Retinopathy, yet standard deep learning models often struggle with high false-negative rates and overfitting on limited medical datasets. Objective: This study proposes a Parallel Comparative Multi-Scenario Framework to identify the most robust configuration for MA detection. The framework evaluates independent 1D vectorized feature descriptors, each initialized as a high-dimensional 16,384-feature baseline, to avoid the redundancy inherent in feature fusion. Methodology: The system systematically processes six independent descriptors LBP, GLCM, Gabor, Wavelet, Fractal, and LMR across three selection tiers (Filter, Wrapper/RFE, and Embedded). These optimized vectors, reduced from the initial 16,384 dimensions to the most discriminative "Best Subsets," serve as uniform inputs for six classifiers: five traditional Machine Learning (ML) models and a proposed representation-consistent 1D-CNN architecture, resulting in 128 experimental scenarios. Results: Experimental evaluation was conducted on a balanced dataset of 740 fundus images derived from two distinct sources: the publicly available MESSIDOR dataset and a clinically acquired dataset from Hospital Universiti Sains Malaysia (HUSM). The model was trained on MESSIDOR data and subsequently evaluated on an independent HUSM test set to assess generalization performance. The results reveal a significant performance gap. The independent LBP-RFE-SVM scenario achieved the highest performance with an accuracy, recall, and precision of 91.00%. In contrast, the best Deep Learning (DL) configuration, Gabor-ANOVA-1DCNN, reached 87.00% accuracy. Notably, while the 1D-CNN maintained a "performance floor" of 60%, ML demonstrated extreme volatility, dropping to 51.00% with global statistical features. The optimal framework significantly minimized the False Negative Rate (FNR) to 6.76%, missing only 5 out of 74 cases.]]>
