BACKGROUND: Entity-level pathologic structures with independent structures and functions are at a mesoscopic scale between the cell-level and slide-level, containing limited structures thus providing fewer instances for multiple instance learning. This restricts the perception of local pathologic features and their relationships, causing semantic ambiguity and inefficiency of entity embedding. METHOD: This study proposes a novel entity-level multiple instance learning. To realize entity-level augmentation, entity component mixup enhances the capture of relationships of contextually localized pathology features. To strengthen the semantic synergy of global and local pathological features, Bayesian collaborative learning is proposed to construct co-optimization of instance and bag embedding. Additionally, pathological prior transfer implement the initial optimization of the global attention pooling thereby fundamentally improving entity embedding. RESULTS: This study constructed a glomerular image dataset containing up to 23 types of lesion patterns. Intensive experiments demonstrate that the proposed framework achieves the best on 19 out of 23 types, with AUC exceeding 90% and 95% on 20 and 11 types, respectively. Moreover, the proposed model achieves up to 18.9% and 14.7% improvements compared to the thumbnail-level and slide-level methods. Ablation study and visualization further reveals this method synergistically strengthens the feature representations under the condition of fewer instances. CONCLUSION: The proposed entity-level multiple instance learning enables accurate recognition of 23 types of lesion patterns, providing an effective tool for mesoscopic histopathology images classification. This proves it is capable of capturing salient pathologic features and contextual relationships from the fewer instances, which can be extended to classify other pathologic entities.