Annotation-free method for breast tumour segmentation in dynamic contrast-enhanced MRI

Abstract

Breast cancer remains the leading cause of cancer-related mortality among women worldwide. Dynamic contrast-enhanced magnetic resonance imaging (DCE-MRI) is a crucial tool for diagnosing breast tumours, as it provides high-resolution images of tissue structures. However, current methods for analysing these scans face two major limitations: they either require labour-intensive manual annotations by radiologists or rely on automated techniques with limited accuracy. To address these challenges, this study proposes an annotation-free method called the ACM (Algorithmic segmentation, Constraint filtering, Model training) that both eliminates the need for manual annotations and increases the segmentation accuracy. Unlike previous methods, which rely solely on deep learning or traditional algorithms, the ACM incorporates strict constraints to filter the pseudo-labels generated by traditional algorithms and integrates these pseudo-labels with unlabelled data for semi-supervised model training. Additionally, novel strategies, including the multi-class strategy and special augmentations, are introduced to mitigate common challenges in AI-based medical image analysis. Experimental validation on a large multicentre dataset comprising 1209 cases demonstrates that our method achieves a Dice similarity coefficient (DSC) of 83.06% in terms of tumour segmentation, approaching the performance of supervised methods. Furthermore, on an external test set, our method attains an Intersection over Union (IoU) of 74.45%, surpassing the best existing unsupervised methods by 23.4%. These results highlight the robustness and effectiveness of the ACM, which has the potential to significantly reduce the workload of radiologists, improve diagnostic consistency, and facilitate earlier breast cancer detection and personalized treatment planning. The source code is publicly available at https://github.com/Ho-Garfield/ACM-pipeline.