Li Zheren, Cui Zhiming, Zhang Lichi, Wang Sheng, Lei Chenjin, Ouyang Xi, Chen Dongdong, Zhao Xiangyu, Liu Chunling, Liu Zaiyi, Gu Yajia, Shen Dinggang, Cheng Jie-Zhi
Shanghai United Imaging Intelligence Co., Ltd., Shanghai 200030, China; The School of Biomedical Engineering, Shanghai Jiao Tong University, Shanghai 200240, China.
The School of Biomedical Engineering, ShanghaiTech University, Shanghai 201210, China.
Comput Biol Med. 2025 Feb;185:109455. doi: 10.1016/j.compbiomed.2024.109455. Epub 2024 Dec 9.
The deep learning technique has been shown to be effectively addressed several image analysis tasks in the computer-aided diagnosis scheme for mammography. The training of an efficacious deep learning model requires large data with diverse styles and qualities. The diversity of data often comes from the use of various scanners of vendors. But, in practice, it is impractical to collect a sufficient amount of diverse data for training. To this end, a novel contrastive learning, MSVCL+, is developed to equip the deep learning models with better style generalizability. Specifically, the multi-style and multi-view unsupervised self-learning scheme is carried out to seek robust feature embedding against style diversity as a pretrained model. Afterward, the pretrained network is further fine-tuned to the downstream tasks, e.g., mass detection, matching, BI-RADS rating, and breast density classification. The proposed method has been evaluated extensively and rigorously with mammograms from various vendor style domains and several public datasets. The experimental results suggest that the proposed domain generalization method can effectively improve performance of four mammographic image tasks on the data from both seen and unseen domains, and outperform many state-of-the-art (SOTA) generalization methods.
深度学习技术已被证明能在乳腺X线摄影的计算机辅助诊断方案中有效解决多个图像分析任务。训练一个有效的深度学习模型需要大量具有不同风格和质量的数据。数据的多样性通常来自使用不同供应商的各种扫描仪。但是，在实践中，收集足够数量的多样化数据用于训练是不切实际的。为此，开发了一种新颖的对比学习方法MSVCL+，以使深度学习模型具有更好的风格通用性。具体而言，执行多风格和多视图无监督自学习方案，以寻找针对风格多样性的鲁棒特征嵌入作为预训练模型。之后，将预训练网络进一步微调至下游任务，例如肿块检测、匹配、BI-RADS分级和乳腺密度分类。所提出的方法已使用来自各种供应商风格域的乳腺X线照片和几个公共数据集进行了广泛而严格的评估。实验结果表明，所提出的域泛化方法可以有效提高在来自已见和未见域的数据上的四个乳腺X线图像任务的性能，并且优于许多最新的（SOTA）泛化方法。
