Accurate preoperative diagnosis of posterior fossa tumors (PFTs) is crucial for treatment and prognosis, yet existing methods fail to sufficiently mine key diagnostic cues from both imaging data and associated reports, resulting in limited robustness. In this paper, we propose Salient Diagnostic Value Perception (SDVP), which integrates MRI images and radiology reports to learn critical diagnostic cues from three complementary perspectives. First, we introduce an Adversarial Intra-sample Contrastive Learning (AICL) method to enhance robustness against variations from different centers and acquisition devices. Second, Knowledge-enhanced Intra-sample Contrastive Learning (KICL) is designed to extract sample-specific diagnostic features under expert guidance. Third, Supervised Inter-class Contrastive Learning (SICL) on clean MRI samples strengthens class-specific feature learning. Additionally, a large …
本文提出显著诊断价值感知方法（SDVP），用于后颅窝肿瘤的术前准确诊断。该方法整合MRI影像与放射学报告，从三个互补视角学习关键诊断线索：通过对抗性样本内对比学习增强跨中心与设备差异的鲁棒性；借助知识增强的样本内对比学习提取专家引导的样本特异性特征；并在干净MRI样本上进行监督式类间对比学习以强化类别特征。