Detailed evaluation of prostate cancer glands is an essential yet labor-intensive step in grading prostate cancer. Gland segmentation can serve as a valuable preliminary step for machine-learning-based downstream tasks, such as Gleason grading, patient classification, cancer biomarker building, and survival analysis. Despite its importance, there is currently a lack of a reliable gland segmentation model for prostate cancer. Without accurate gland segmentation, researchers rely on cell-level or human-annotated regions of interest for pathomic and deep feature extraction. This approach is sub-optimal, as the extracted features are not explicitly tailored to gland information. Although foundational segmentation models have gained a lot of interest, we demonstrated the limitations of this approach. This work proposes a prostate gland segmentation framework that utilizes a dual-path Swin Transformer UNet structure and leverages Masked Image Modeling for large-scale self-supervised pretaining. A tumor-guided self-distillation step further fused the binary tumor labels of each patch to the encoder to ensure the encoders are suitable for the gland segmentation step. We united heterogeneous data sources for self-supervised training, including biopsy and surgical specimens, to reflect the diversity of benign and cancerous pathology features. We evaluated the segmentation performance on two publicly available prostate cancer datasets. We achieved state-of-the-art segmentation performance with a test mDice of 0.947 on the PANDA dataset and a test mDice of 0.664 on the SICAPv2 dataset.
对前列腺癌腺体进行详细评估是前列腺癌分级中至关重要但劳动密集的步骤。腺体分割可作为基于机器学习的下游任务（如格里森分级、患者分类、癌症生物标志物构建和生存分析）的重要预处理步骤。尽管其重要性不言而喻，目前仍缺乏可靠的前列腺癌腺体分割模型。在缺乏准确腺体分割的情况下，研究者通常依赖细胞层面或人工标注的感兴趣区域进行病理组学和深度特征提取。这种方法并非最优解，因为提取的特征并未明确针对腺体信息进行优化。尽管基础分割模型已获得广泛关注，我们论证了该方法的局限性。本研究提出一种前列腺腺体分割框架，采用双路径Swin Transformer UNet结构，并利用掩码图像建模进行大规模自监督预训练。通过肿瘤引导的自蒸馏步骤，将每个图像块的二元肿瘤标签融合至编码器，确保编码器适用于腺体分割任务。我们整合了活检与手术标本等异质性数据源进行自监督训练，以涵盖良恶性病理特征的多样性。在两个公开的前列腺癌数据集上评估了分割性能：在PANDA数据集上获得0.947的测试mDice分数，在SICAPv2数据集上获得0.664的测试mDice分数，均达到当前最优分割性能。
肿瘤(癌症)患者之家