Few-shot learning with deformable convolution for multiscale lesion detection in mammography

Image-based breast lesion detection is a powerful clinical diagnosis technology. In recent years, deep learning architectures have achieved considerable success in medical image analysis however, they always require large-scale samples. In mammography images, breast lesions are inconspicuous, multiscale, and have blurred edges. Moreover, few well-labeled images exist. Because of these factors, the detection accuracy of conventional deep learning methods is low. Therefore, we attempted to improve the accuracy of mammary lesion detection by introducing transfer learning (TL) into a deep learning framework for the few-shot learning task and thus provide a method that will further assist physicians in detecting breast lesions.

In this paper, we propose a method called "few-shot learning with deformable convolution for multiscale lesion detection in mammography," named FDMNet. Deformable convolution is introduced for enhancing the network's ability to detect lesions, and the sensitivity of the multiscale feature space is reinforced by using a feature pyramid method. Furthermore, by introducing location information in the predictor, the sensitivity of the model to lesion location is also enhanced. The proposed method, through the TL technique that is applied mines the potentially common knowledge of features in the source domain and transfers it into the target domain to improve the accuracy of breast lesion detection in the few-shot learning task.

On the publicly available datasets for screening mammography CBIS-DDSM and Mini-MIAS, the proposed method performs better than five widely used detection methods. On the CBIS-DDSM dataset, its comprehensive scores, sensitivity, precision, and the mean dice similarity coefficient are 0.911, 0.949, 0.873, and 0.913, respectively, and on the Mini-MIAS dataset, these values are 0.931, 0.966, 0.882, and 0.941, respectively.

To achieve the few-shot learning required for medical image analysis, the proposed method uses TL to execute feature knowledge transformation and includes deformable convolution to build a feature pyramid structure, which enhances the learning performance of the network for lesions. The results of comparative numerical experiments show that the proposed method outperforms some state-of-the-art methods.