A LIGHTWEIGHT DEEP LEARNING-BASED CERVICAL IMAGE RECONSTRUCTION FRAMEWORK TO SUPPORT RELIABLE MOBILE HEALTH DIAGNOSTICS IN RESOURCE-CONSTRAINED ENVIRONMENTS

ABSTRACT

This study presents a lightweight deep neural network–based framework for cervical image reconstruction to improve the reliability of mobile health (mHealth) image transmission under adverse wireless conditions. Using the AVIVA mobile application as a real-world case study, the proposed framework addresses major issues that are common in low-resource clinical settings, such as low-resolution image capture, wireless channel impairments, and limited computing power. A dataset of 3,000 high quality images of the cervix was used, and realistic degradations were added to simulate Rician fading effects that are normally experienced when transmitting mobile images. The model involves the application of deep convolutional autoencoder that incorporates skip connections and structural similarity-based optimization to recover degraded images without affecting key diagnostic features. Results show strong performance, achieving SSIM of 0.9424, PSNR of 35.83 dB, SNR of 31.73 dB, and BER of 0.0168, with low latency (0.1571 s/image) and modest computational complexity (683,779 parameters). The model demonstrates efficiency in reliable, real-time mHealth implementation in resource constrained environments.

Keywords: Deep Learning, Medical Imaging, Image Reconstruction, mHealth, Wireless Communication