Breast cancer is the most common cancer among women, and in some cases, it also affects men. Since early detection allows for proper treatment, automated data classification is essential. Although such classifications provide timely results, the resource requirements for such models, i.e., computation and storage, are high. As a result, these models are not suitable for resource-constrained devices (for example, IOT). In this work, we highlight the U-Net model, and to deploy it to IOT devices, we compress the same model using a genetic algorithm. We assess the proposed method using a publicly accessible, bench-marked dataset. To verify the efficacy of the suggested methodology, we conducted experiments on two more datasets, specifically CamVid and Potato leaf disease. In addition, we used the suggested method to shrink the MiniSegNet and FCN 32 models, which shows that the compressed U-Net approach works for classifying breast cancer. The results of the study indicate a significant decrease in the storage capacity of UNet with 96.12% compression for the breast cancer dataset with 1.97x enhancement in inference time. However, after compression of the model, there is a drop in accuracy of only 1.33%.
Keywords: Acceleration; Compression; Deep learning; FCN; Genetic algorithm; MiniSegNet; Semantic segmentation; UNet.
Copyright © 2024 Elsevier Ltd. All rights reserved.