Reconstruction-based anomaly detection models are widely used in medical imaging for their ability to provide both image- and pixel-level anomaly estimations, enhancing interpretability. However, in complex medical scenarios, these models often capture generic features, leading to the unintended reconstruction of anomalies and ultimately failing in the anomaly detection task. In this study, we investigate the application of reconstruction-based methods to anterior eye segment (AES) anomaly detection and propose a novel training framework based on Self-Supervised (SSL) Siamese Autoencoders to enhance normality representation learning. We introduce a combined self-distillation and restoration objective using pseudo-anomalies to mitigate overgeneralization, while leveraging perceptual and structural reconstruction losses for improved reconstruction quality. We evaluated our model on a private dataset of 2,647 external photographs of the AES - including clinical samples and 36 identified diseases such as cataracts, hemorrhages, pinguecula, and pterygium - and on a public dataset of 3,200 retinal fundus images with 46 annotated pathologies. Experimental results demonstrate superior quantitative and qualitative performance compared to baseline methods on both datasets. Our approach shows strong generalization capabilities and holds promise for non-invasive AES biomarker discovery. This work marks, to the best of our knowledge, the first attempt to apply unsupervised anomaly detection models to external ocular photographs, a non-invasive and cost-effective ophthalmic imaging modality. Code is available at: https://github.com/VisionICLab/AnoSiamAE .