High annotation costs and one-sided label noise (benign samples mislabeled as malignant) pose challenges in digital pathology. To address these issues, a One-Sided Label Noise Multi-Instance Learning (MIL-OSLN) framework based on a dual-weighting mechanism was proposed to improve the automatic diagnosis accuracy of ovarian cancer pathological images. The framework dynamically weighted samples at both instance and bag levels simultaneously. It synergistically screened key diagnostic regions from noisy data and suppressed interference from false-positive bags. The model was trained using a weak label dataset generated via semi-supervised clustering, and performance was compared against various baseline models on an independent gold-standard test set. Experimental results indicated that the model"s Area Under the Curve (AUC) remained stable at approximately 0.9 even when the training set noise rate reached 50%. The comprehensive performance significantly outperformed all comparative baseline models. The study demonstrates that this method effectively addresses one-sided label noise in MIL and is suitable for intelligent tumor diagnosis using large-scale weakly labeled pathological data.