Abstract
Automated 3D chest CT report generation is domi- nated by a single-report paradigm: one textual report per input volume via greedy or beam decoding. Under the standard clini- cal-efficacy (CE) F1 evaluation over the 18 chest-CT pathologies extracted by a pre-trained RadBERT clinical extractor, this paradigm shows high precision but low recall, omitting many pathologies. Motivated by this recall bottleneck, we propose MDEF, a deep-ensemble pipeline that turns the single-report paradigm into multi-report paradigm. The multi-report stages contain three fine-tuned report generators with Stochastic De- coding per generator, so that each generator samples different reports for each input volume. Driven by the deep-ensemble as- sumption that the ensemble’s joint per-pathology miss probability is smaller than that of any individual generator, this multi-report design lowers the per-pathology miss probability and therefore raises per-pathology recall. To consolidate all sampled reports into a single final report that preserves every pathology found by any of them, each generator first concatenates its K samples into one extended report, the RadBERT extractor then maps each generator’s extended report to an 18-pathology binary vector, and a logical-OR aggregator fuses those per-generator vectors into the final 18-pathology binary vector. Hence, any pathology mentioned in any sampled report is retained in the final prediction. Lastly, a deterministic text head generates the final report from the final 18-pathology binary vector and the per-generator extended reports. Trained only on RadGenome and evaluated on RadGenome, CT-RATE, and INSPECT, our MDEF attains CE F1 of 0.3999, 0.3381, and 0.1813 respec- tively. It outperforms the state-of-the-art single-report baselines we evaluate by clear margins on RadGenome and CT-RATE, and by a small positive margin on INSPECT. The CE F1 gain is driven by a large recall improvement. Averaged over the 18 pathologies, recall rises by +70% to +97% relative across the three datasets. Our source code is available at: https://anonymous.4open.science/r/mdef-icdm2026-42B1.