Systematic Evaluation of Autoencoder Architectures for Medical Image Reconstruction: A Comprehensive Study Across Six...

Autoencoder-based representational learning models have become extremely popular tools enabling the analysis of medical images, yet systematic comparisons across diverse medical imaging modalities remain limited. Understanding which autoencoder architectures perform optimally for specific medical imaging tasks is crucial for advancing healthcare applications. This study aims to conduct a comprehensive systematic comparison of six autoencoder architectures across all available medical imaging datasets from MedicalMNIST, while introducing novel evaluation metrics to better assess both reconstruction quality and latent space organization. We evaluated Vanilla Autoencoders, Variational Autoencoders, Beta-Variational Autoencoders, Convolutional Autoencoders, Denoising Autoencoders, and Sparse Autoencoders across two-dimensional medical image modalities. Two novel metrics were introduced: the Reconstruction Quality Score (RQS), capturing pixel-level fidelity and structural similarity, and the Latent Space Quality Index (LSQI), assessing organization, separability, and coverage of learned representations. Experiments encompassed the complete MedicalMNIST collection, covering pathology, radiology, dermatology, ophthalmology, and organ datasets. Our results demonstrate that the Convolutional Autoencoder consistently outperforms other variants across all 12 modalities, achieving an average RQS of 0.941 and LSQI of 0.726, representing a 15.4% improvement in structural fidelity over the Vanilla AE baseline. While Variational and Sparse models showed marginal reconstruction trade-offs, they maintained competitive latent organization, with VAEs achieving a mean Pixel Correlation of 0.931 in radiological datasets. This comprehensive evaluation provides definitive guidance for autoencoder selection in medical imaging applications. The introduced RQS and LSQI metrics offer deeper insights into autoencoder performance beyond conventional error-based measures, addressing critical limitations in current evaluation approaches for medical image analysis.