Recognition of breast cancer from heterogeneous ultrasound images: A multi-level deep learning approach

Breast ultrasound is a medical imaging technique that employs sound waves to produce breast images, and it has been primarily used to diagnose breast cancer and other related issues. With various machine learning algorithms being applied, many applications have shown promising results and demonstrated outstanding efficiency in giving doctors early accurate diagnoses. By investigating existing state-of-the-art approaches to breast lesion detection, given ConvNeXt-Small architecture as an example, we observe that although they bring a satisfactory performance in classification, their ability to detect small lesions is limited. Therefore, there is still room for improving the performance of DL-based approaches. In this paper, we present a practical Deep Learning-based solution for breast lesion detection, using DetectoRS with Gaussian Receptive Field-based Label Assignment (RFLA) and SegFormer-B4 for recognizing and segmenting small breast lesions, including malignant tumors. Our proposed solution involves three distinct models: ConvNeXt-Small for classification, Swin-Base combined with DetectoRS and RFLA for object detection, and SegFormer-B4 for segmentation. Each model is tailored specifically to address its respective task in breast cancer detection and analysis. The proposed approach has been evaluated on diverse ultrasound datasets. Our deep learning model achieves an Average Precision of 0.270 for small objects (AP_S), and records the highest mean Intersection over Union at 81.55%. The results show that the proposed model outperforms various well-established baselines. We suppose that our method can be integrated into computer-aided diagnosis systems to assist physicians in their clinical activities.