Efficient Feature Learning–Driven Neuroimage Compression and High-Fidelity Reconstruction using a Lightweight Autoencoder

Medical neuroimaging, and particularly MRI and CT, forms the cornerstone of current diagnosis and treatment planning owing to its accurate structural information of the brain. But the exponentially increased size and resolution of neuroimaging data significantly challenge its storage and retrieval. Traditional lossy compression schemes may affect the diagnostic features present in the medical imaging, and they rely on pre-specified transformation rather than adaptive feature learning. To address this, a lightweight SegNet and UNet++-inspired autoencoder, Medical Neuroimaging Lightweight Autoencoder (MedLiteNeuro-AE) is proposed in this research work that can compress the brain MRI and CT scan images before storing hospital environments. The proposed model compresses images into the smallest latent representation and reconstructs them by preserving diagnostically reliable fidelity. Unlike conventional codecs, the latent code layer also acts as a non-interpretable compressed representation, reducing direct visual interpretability of the original scans during storage and transmission, while clinicians interact only with the reconstructed images. Extensive evaluation proves the effectiveness of MedLiteNeuro-AE, with regeneration metrics of 3.1E-05 MSE, 5.042E-03 RMSE, 46.54 dB PSNR, 99.66% SSIM, and 99.66% MS-SSIM, along with a compression rate of 3.12% and model size of 3.87 MB. The results confirm that our model achieves a balance among the effective compression, preservation of diagnostic quality, and computational complexity, and outperforms common compression baselines while being significantly. Later, the MedLiteNeuro-AE can be decomposed into an encoder and a decoder to build a feature learning-based compressor and decompressor. These results position MedLiteNeuro-AE as a safe, lightweight, and adaptable solution for modern neuroimaging deployments.