DDSNet: A Lightweight Dense Depthwise Separable Network for Tumor Classification

An Huang, Junggab Son, Zuobin Xiong

Abstract: Deep learning-based medical image processing plays a significant role in modern computer-aided diagnosis, which facilitates doctors in various disease analysis. However, most researchers focus on the accuracy of medical image classification tasks with ever-increasing model size and the number of parameters but overlook the high diagnostic costs and model efficiency. To reduce such costs and broaden the application scenarios, a low-cost and efficient medical image classification is imperative. To achieve this goal, this paper designs a lightweight model, named Dense Depthwise Separable Network (DDSNet), which combines the merits of Dense Convolution Network and Depthwise Separable Convolution, rendering a low-cost and efficient medical imaging. Moreover, a quantization-based method is invented to deploy the proposed model on real-world IoT devices by converting the original model to an integer-type model while maintaining its classification performance. Extensive experiments are conducted on four cancer image datasets on the IoT device, showing the promising performance of this proposed method against 5 baseline models, including data visualization and interoperability aspects. Notably, compared to DenseNet, the proposed model is about 32× smaller and 5× faster after quantization, with a competitive classification accuracy preserved. Our code is available at https://github.com/OldDreamInWind/DDSNet.

External IDs:dblp:conf/sac/HuangSX25