"Accelerating Deep Learning for Medical Image Analysis"

"The processing of large-scale, high-resolution medical images often leads to substantial computational requirements, increased GPU memory consumption, and longer training durations. To overcome these challenges, this work proposes an optimized deep learning framework for diabetic retinopathy prediction, integrating three key optimization strategies within a unified pipeline: mixed-precision training to reduce memory usage, Accelerated Linear Algebra (XLA) compilation to improve execution efficiency through operation fusion, and an asynchronous data pipeline to enhance data input throughput. The proposed model utilizes a lightweight convolutional neural network architecture incorporating batch normalization, ReLU activation functions, and dropout layers to enable stable and effective feature extraction. The input data consist of Gaussian-filtered retinal fundus images, which enhance relevant visual patterns and assist in accurate detection of diabetic retinopathy."