CascadeNet: A Multi-Stage Deep Learning Framework for Breast Cancer Detection in Ultrasound Imaging

Breast cancer is still the most common cancer detected in women around the globe. The early and accurate detection of breast cancer is essential to improve outcomes in patients; however, these can be problematic with current modalities, especially if the tumour is obscured by dense breast tissue. The use of ultrasound imaging can fulfil this need with a powerful alternative due to its ability to capture soft tissue anatomy. In this article, a novel cascaded multistage deep learning architecture that integrates YOLOv12 for tumor localization and a modified Detection Transformer (DETR) without object queries for tumor classification is proposed. The model benefits from the accuracy of detection with YOLOv12 coupled with the transformer classification with the DETR model. An experimental evaluation indicates the mAP for the model (mAP@50) improved (96.41%), recall (90.37%), and F1 score (93.64%) compared to the single models. Most notably, the model is in real-time on a clinical embedded system. Future work will be directed towards improvements with semi-supervised learning approaches and expand the model study to enhance overall diagnostic accuracy.