This study aims to investigate the effectiveness of self-knowledge distillation (self-KD) with progressive refinement in the early prediction of molecular subtypes of breast cancer (BC) using digital breast tomosynthesis (DBT) images. This study conducted a retrospective analysis of 368 patients who underwent breast DBT and/or magnetic resonance imaging (MRI) scans at our hospital. Among these patients, 303 underwent DBT scans and 119 underwent MRI scans. Of the DBT patients, 137 had images with molecular subtypes labels, while the remaining 166 did not have molecular subtype annotations. None of the MRI patients had the corresponding molecular subtype labels. To address the issue of insufficient labeled DBT images, we proposed a self-knowledge distillation (self-KD) framework with progressive refinement to more effectively utilize the unlabeled MRI and DBT image. Initially, the teacher network was pre-trained using unlabeled MRI images to capture the essential characteristics of BC. Subsequently, the teacher network was progressively refined to generate more accurate soft labels for the unlabeled DBT images, which improved the performance of the student network through KD. Additionally, a noise-adaptive layer was integrated to adjust the soft labels for more accurate learning. The performance of our method was evaluated using the area under the receiver operating characteristic curve (AUC), accuracy (ACC), sensitivity (SEN), and specificity (SPE) values. The proposed self-KD method achieved an AUC of 0.834, ACC of 0.732, SEN of 0.930, and SPE of 0.734, which surpassed the existing methods for BC molecular subtype prediction. Specifically, compared to the baseline KD, our self-KD improved AUC by 9%, ACC by 6%, SEN by 26%, and SPE by 9%. The proposed self-KD framework effectively refines the network using both labeled and unlabeled images, which enables more accurate BC molecular subtype prediction.

Keywords: Breast cancer; DBT; MRI; Molecular subtype; Self-knowledge distillation.

© 2025. International Federation for Medical and Biological Engineering.

本研究旨在通过数字乳腺断层合成（DBT）图像，探究自我知识蒸馏（self-KD）结合渐进式改进在早期预测乳腺癌（BC）分子亚型中的有效性。本研究对368名在我院接受过乳腺DBT和/或磁共振成像（MRI）检查的患者进行了回顾性分析。在这368名患者中，有303人接受了DBT扫描，119人接受了MRI扫描。在DBT组中，137人的图像带有分子亚型标签，其余166人没有相应的分子亚型注释。而所有接受MRI检查的患者都没有对应的分子亚型标签。

为了解决标注不足的DBT图像问题，我们提出了一种结合渐进式改进的自我知识蒸馏（self-KD）框架，以更有效地利用未标记的MRI和DBT图像。首先，使用未标记的MRI图像预训练教师网络，捕捉BC的基本特征。随后，该教师网络逐步优化，为未标注的DBT图像生成更为准确的软标签，进而通过知识蒸馏（KD）提高学生网络的表现力。此外，还集成了噪声自适应层来调整软标签以实现更精确的学习。

我们的方法使用受试者工作特性曲线下的面积（AUC）、准确性（ACC）、敏感性（SEN）和特异性（SPE）值进行了评估。所提出的self-KD方法达到了0.834的AUC、0.732的ACC、0.930的SEN以及0.734的SPE，超越了现有的BC分子亚型预测方法。特别是与基线KD相比，我们的self-KD使AUC提高了9%，ACC提高了6%，SEN提高了26%，SPE提高了9%。

所提出的self-KD框架能够有效地利用标记和未标记图像对网络进行精炼，从而实现更准确的BC分子亚型预测。 关键词：乳腺癌；DBT；MRI；分子亚型；自我知识蒸馏。