腹膜后脂肪肉瘤（RLPS）是一种罕见的恶性肿瘤，其唯一的治疗方法是手术切除。然而，高分化脂肪肉瘤（WDLPS）是其最常见的类型之一，如果没有有效的切缘评估方法，在手术过程中很难将其与正常脂肪区分开，严重危害预后且复发风险很高。在这里，我们结合了双无标记非线性光学模式、受激拉曼散射 (SRS) 显微镜和二次谐波发生 (SHG) 显微镜，对 35 个 RLPS 和 34 个正常脂肪样本中的两种主要组织生物分子、脂质和胶原纤维进行成像。 35名患者。生成的双模态组织图像用于基于深度学习的 RLPS 诊断。反映了肿瘤进展过程中脂质的显着降低和胶原纤维的增加。基于 ResNeXt101 的模型在区分正常脂肪、WDLPS 和去分化脂肪肉瘤 (DDLPS) 方面实现了 94.7% 的总体准确率和 0.987 ROC 曲线下平均面积 (AUC)。特别是，WDLPS 的检测精度和灵敏度分别为 94.1% 和 84.6%，优于现有方法。消融实验表明，这种性能归功于SRS和SHG显微镜，它们分别将识别WDLPS的灵敏度提高了16.0和3.6%。此外，我们在 RLPS 边缘上利用该模型来识别肿瘤浸润。我们的方法对于术中脂肪肉瘤的准确检测具有巨大的潜力。

The retroperitoneal liposarcoma (RLPS) is a rare malignancy whose only curative therapy is surgical resection. However, well-differentiated liposarcomas (WDLPSs), one of its most common types, can hardly be distinguished from normal fat during operation without an effective margin assessment method, jeopardizing the prognosis severely with a high recurrence risk. Here, we combined dual label-free nonlinear optical modalities, stimulated Raman scattering (SRS) microscopy and second harmonic generation (SHG) microscopy, to image two predominant tissue biomolecules, lipids and collagen fibers, in 35 RLPSs and 34 normal fat samples collected from 35 patients. The produced dual-modal tissue images were used for RLPS diagnosis based on deep learning. Dramatically decreasing lipids and increasing collagen fibers during tumor progression were reflected. A ResNeXt101-based model achieved 94.7% overall accuracy and 0.987 mean area under the ROC curve (AUC) in differentiating among normal fat, WDLPSs, and dedifferentiated liposarcomas (DDLPSs). In particular, WDLPSs were detected with 94.1% precision and 84.6% sensitivity superior to existing methods. The ablation experiment showed that such performance was attributed to both SRS and SHG microscopies, which increased the sensitivity of recognizing WDLPS by 16.0 and 3.6%, respectively. Furthermore, we utilized this model on RLPS margins to identify the tumor infiltration. Our method holds great potential for accurate intraoperative liposarcoma detection.