Low-Power Quantized Convolutional Neural Network for Early Breast Cancer Detection in Remote Communities

Breast cancer remains a leading cause of mortality in low-resource communities due to late diagnosis and limited access to specialized healthcare facilities. Leveraging recent advancements in deep learning, this study presents a low-power, cost-effective solution for early breast cancer detection tailored to resource-constrained environments. We developed a quantized convolutional neural network (CNN) using Brevitas and deployed it on a PYNQ-Z1 development board via the FINN framework. The CNN efficiently classifies ultrasonic topography images into three categories: normal, benign, or malignant. Achieving up to 91.2% accuracy on the BUSI dataset, our results highlight the effectiveness of 4-bit quantization, offering a viable trade-off between computational efficiency and accuracy. This makes it viable for real-time medical image classification in underserved communities, potentially facilitating early diagnosis and timely referrals, and ultimately contributing to improved healthcare outcomes.