Entity-level multiple instance learning for mesoscopic histopathology images classification with Bayesian...

Area of Science:

• Histopathology
• Computational Pathology
• Machine Learning

Background:

• Pathologic structures exist at a mesoscopic scale, posing challenges for traditional multiple instance learning due to limited instances.
• This limitation hinders the perception of local features and their relationships, leading to semantic ambiguity and inefficient entity embedding.

Purpose of the Study:

• To develop a novel entity-level multiple instance learning framework for improved mesoscopic histopathology image classification.
• To address the challenges of limited instances and semantic ambiguity in entity embedding.

Main Methods:

• Proposed a novel entity-level multiple instance learning approach.
• Implemented entity component mixup for enhanced capture of localized pathology features.
• Utilized Bayesian collaborative learning for co-optimization of instance and bag embedding.
• Applied pathological prior transfer for initial optimization of global attention pooling.

Main Results:

• Achieved state-of-the-art performance on 19 out of 23 lesion types in a glomerular image dataset.
• Demonstrated AUC exceeding 90% on 20 types and 95% on 11 types.
• Showcased significant improvements (up to 18.9% and 14.7%) over thumbnail-level and slide-level methods.
• Ablation studies confirmed synergistic strengthening of feature representations with fewer instances.

Conclusions:

• The proposed entity-level multiple instance learning framework enables accurate classification of 23 lesion patterns in mesoscopic histopathology images.
• The method effectively captures salient pathologic features and contextual relationships from limited instances.
• This approach offers a promising tool for histopathology image classification and can be extended to other pathologic entities.