Genetic Algorithm-Based Optimization of UNet for Breast Cancer Classification: A Lightweight and Efficient Approach for IoT Devices

IoT devices are widely used in medical domain for detection of high blood sugar and life threatening disease such as cancer. Breast cancer is one of the most challenging type of cancer which not only affects women but in some cases men also. Deep learning is one of the widely used technology which provides efficient classification of cancerous lumps but it is not useful for IoT devices as the devices lack resources such as storage and computation. For the suitability in IoT devices, in this work, we are compressing UNet, the popular semantic segmentation technique, for the pixel-wise classification of breast cancer. For compressing the deep learning model, we use genetic algorithm which removes the unwanted layers and hidden units in the existing UNet model. We have evaluated the proposed model and compared with the existing model(s) and found that the proposed compression technique suppresses the storage requirement to 77.1%. Additionally, it also improves the inference time by 3.82 $$\times $$ without compromising the accuracy. We conclude that the primary reason of inference time improvement is the requirement of less number of weight and bias by the proposed model.