MDEF: Multi-Report Deep Ensemble Fusion for 3D CT Report Generation

[NEW PUBLICATION]

Congratulations to the author group Tung Duong Ta, Tim Oates, Manas Gaur, Trong-Nghia Nguyen, Tien-Cuong Nguyen, Tuan-Cuong Vuong, Trang Xuan Mai and Thien Van Luong for having their research paper officially accepted at the ICDM 2026 International Conference - an A* rated conference.

• Conferences:ICDM 2026 (International Conference on Data Mining)

• Rank:A*

• Article title:MDEF: Multi-Report Deep Ensemble Fusion for 3D CT Report Generation

Main contributions:

• Proposing a pioneering MDEF architecture: Building a deep ensemble process consisting of 5 stages, combining Stochastic Decoding and logical-OR late fusion mechanism, to overcome the limitation of missing pathologies of the single-report paradigm in the problem of generating 3D chest CT reports.

• Improved clinical coverage: Improved ability to detect 18 pathologies according to RadBERT, with average recall increased from 70% to 97% compared to modern single models such as MARCH, Reg2RG and CT-GRAPH.

• Ability to generalize across multiple data distributions: Achieve stable clinical F1 (CE F1) scores on both same-source and cross-source data from hospitals and imaging devices, with RadGenome: 0.3999, CT-RATE: 0.3381 and INSPECT: 0.1813 results with just a single configuration.

This achievement contributes to affirming the potential of advanced AI methods in improving the quality of automatically generating diagnostic imaging reports, supporting the process of analyzing and detecting pathology in clinical practice.