Data pyramid structure for optimizing EUS-based GISTs diagnosis in multi-center analysis with missing label.

Area of Science:

• Medical Image Analysis
• Machine Learning in Healthcare
• Computational Pathology

Background:

• Data sparsity and missing labels are significant challenges in medical image analysis.
• Existing methods struggle with multi-center data integration and incremental learning, especially with incomplete datasets.
• Accurate attribute prediction and malignant tumor diagnosis are critical for effective patient care.

Purpose of the Study:

• To introduce the Data Pyramid Structure (DPS) to overcome data sparsity and missing labels in medical image analysis.
• To optimize multi-task learning for attribute prediction and malignant tumor diagnosis.
• To enable sustainable expansion of multi-center data analysis using novel frameworks.

Main Methods:

• Developed the Data Pyramid Structure (DPS) for segmentation and aggregation of data with absent attribute labels.
• Proposed the Unified Ensemble Learning Framework (UELF) and Unified Federated Learning Framework (UFLF) for multi-center data leveraging.
• Incorporated strategies for data transfer and incremental learning in missing label scenarios.

Main Results:

• Evaluated the proposed method on a challenging EUS patient dataset from five centers.
• Achieved an average accuracy of 0.984 and an AUC of 0.927 for multi-center analysis.
• Demonstrated superior performance compared to state-of-the-art approaches, with interpretable predictions.

Conclusions:

• The Data Pyramid Structure (DPS) effectively addresses data sparsity and missing labels in medical imaging.
• The UELF and UFLF frameworks facilitate robust multi-center data analysis and incremental learning.
• The method shows significant potential for clinical application in malignant tumor diagnosis.