Task-specific neural networks for medical imaging using pretrained fragments

The StitchNet framework introduced a paradigm shift in Neural Architecture Search (NAS) by proposing the construction of neural networks from pre-trained fragments. This approach reduces computational costs and enables task-specific model creation without retraining entire networks. Building on this foundation, our study evaluates the practical application of StitchNet in constructing neural networks tailored to medical image classification tasks. Specifically, we assess its performance on a dataset of retinal images classified into three categories: healthy, dry, and wet AMD (Age-Related Macular Degeneration), namely drusen and choroidal neovascularization (CNV). By employing fragments from five pre-trained networks and integrating techniques such as recurrent neural networks (RNNs) and autoencoders, we aim to validate and enhance StitchNet's capabilities. Our findings demonstrate that while StitchNet achieves competitive accuracy with reduced computational overhead, incorporating domain-specific optimizations further improves its adaptability and efficiency. So, the developed network outperforms a scientist-designed network by 6%. In the next phase, we will explore ways to improve the algorithm's efficiency and minimize the data required for processing. Fully reproducible code here: https://github.com/ShafighAshrafi/stitchnet.