Ovarian cancer pathological tissues exhibit high morphological heterogeneity and complex microenvironmental features, with benign and malignant regions often showing subtle and blurred boundaries visually, posing significant challenges to pathological diagnosis. High annotation costs and one-sided label noise, where benign samples are mislabeled as malignant, pose major challenges in digital pathology. To address these issues, this study proposes a one-sided label noise multi-instance learning (MIL-OSLN) framework based on a dual-weighting mechanism to improve the automated diagnostic accuracy of ovarian cancer pathological images. The proposed framework dynamically assigns weights at both the instance and bag levels. This strategy enables the model to identify diagnostically relevant regions from noisy data while suppressing interference from false-positive bags. The model was trained on a weakly labeled dataset generated via semi-supervised clustering, and its performance was evaluated against multiple baseline models using an independent gold-standard test set. Experimental results indicated that the model maintained a stable area under the curve (AUC) of approximately 0.9 even when the training set noise rate reached 50%. Its overall performance significantly outperformed all comparative baseline models. These findings demonstrate that the proposed method effectively addresses one-sided label noise in multi-instance learning and is well suited for intelligent tumor diagnosis using large-scale weakly labeled pathological data.