Deep Learning Approaches for Simultaneous Cataract and Anemia Detection from Ocular Images Using EfficientNet-B0 and MobileNetV3-Small

This paper presented the design, training, and deployment of a dual-pipeline ocular screening system pairing EfficientNet-B0 for cataract classification with MobileNetV3-Small for conjunctival anemia detection. The anemia module achieved 87.50% accuracy and an AUC of 0.9023 on a held-out test partition, with zero false positives recorded — a performance profile consistent with published single-disease anemia classifiers and sufficient to justify continued development toward clinical validation. The cataract module achieved 83.46% validation accuracy, an AUC of 0.8992, and a recall of 0.9768 on the independent test partition, with a generalization gap to 56.25% test accuracy attributed to dataset-acquisition heterogeneity and test partition size rather than a fundamental model failure. Both pipelines are integrated within a fully operational Flask web application that executes sub-second inference, applies confidence-threshold gating, and generates structured PDF screening reports. Planned future work addresses the three most significant limitations identified in this study. Dataset expansion is the highest priority: enlarging both the anemia-positive and normal-lens subsets, with images collected across multiple devices, lighting environments, and demographic groups, is prerequisite to generating reliable generalization estimates. Grad-CAM visualization will be integrated to confirm that model attention regions correspond to clinically meaningful anatomical structures. Model compression through quantization and pruning will be evaluated to support inference on mobile hardware. Finally, the evaluation framework will be upgraded to k-fold cross-validation and a minimum 200-image per-class held-out test set, enabling statistically defensible performance claims for subsequent submissions.

[1]     P. Appiahene, Y. B. Missah, U. Danyuo, and E. K. Donkoh, “Detecting Anemia in Childhood Using Conjunctival Image Processing and Machine Learning Methods,” IEEE Access, vol. 10, pp. 45191–45205, 2022. [2]  J. Lee, S. Kim, H. Park, and Y. Lee, “Deep learning-based non-invasive anemia detection from palpebral conjunctival images under ambient lighting conditions,” npj Digital Medicine, vol. 8, no. 1, pp. 1–10, 2025. [3]  A. S. Razavian, H. Azizpour, J. Sullivan, and S. Carlsson, “CNN features off-the-shelf: An astounding baseline for recognition,” in Proc. IEEE CVPR Workshops, 2014, pp. 806–813. [4]   S. Dhalla, S. Sheth, K. Shah, and J. Patel, “Conjunctival pallor detection using deep segmentation and classification networks for non-invasive anemia screening,” in Proc. IEEE EMBC, 2022, pp. 1894–1897. [5]   V. Gulshan et al., “Development and validation of a deep learning algorithm for detection of diabetic retinopathy in retinal fundus photographs,” JAMA, vol. 316, no. 22, pp. 2402–2410, 2016. [6]  A. Esteva et al., “Dermatologist-level classification of skin cancer with deep neural networks,” Nature, vol. 542, pp. 115–118, 2017. [7]   S. Devaraj, G. Raman, R. Shankar, and A. Kumar, “AI-based mobile cataract screening in rural India: Field deployment across 14 districts of Uttar Pradesh,” Lancet Digit. Health, vol. 5, no. 3, pp. e171–e179, 2023. [8]   P. Appiahene, J. Ben Hayfron-Acquah, I. Asante, and T. Frempong, “Smartphone application for real-time non-invasive anemia detection using deep learning on conjunctival images,” Diagnostics, vol. 13, no. 15, Art. no. 2574, 2023. [9]     M. Tan and Q. Le, “EfficientNet: Rethinking model scaling for convolutional neural networks,” in Proc. ICML, 2019, pp. 6105–6114. [10]  A. Howard et al., “Searching for MobileNetV3,” in Proc. ICCV, 2019, pp. 1314–1324.