Bootstrapping Large Language Models for Radiology Report Generation

Authors: Chang Liu, Yuanhe Tian, Weidong Chen, Yan Song, Yongdong Zhang

AAAI 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Experimental results on two prevailing RRG datasets, namely, IU X-Ray and MIMIC-CXR, demonstrate the superiority of our approach to previous state-of-the-art solutions.
Researcher Affiliation Academia Chang Liu1, Yuanhe Tian2, Weidong Chen1, Yan Song1 , Yongdong Zhang1 1University of Science and Technology of China 2University of Washington lc980413@mail.ustc.edu.cn, yhtian@uw.edu, chenweidong@ustc.edu.cn, clksong@gmail.com, zhyd73@ustc.edu.cn
Pseudocode No The paper describes the proposed approach in narrative text and with a system architecture diagram, but it does not include pseudocode or an algorithm block.
Open Source Code Yes Code and models of our approach are available at https:// github.com/synlp/R2-LLM.
Open Datasets Yes We conduct our experiments on two conventional benchmark datasets, i.e., IU X-RAY (Demner-Fushman et al. 2016) from Indiana University and MIMIC-CXR (Johnson et al. 2019) from the Beth Israel Deaconess Medical Center.
Dataset Splits Yes We follow the dataset split in Li et al. (2018) for IU X-RAY and the official split of MIMIC-CXR. Table 1 reports the statistics of all datasets in terms of the numbers of radiographs, reports, patients, and average report length according to each split of the datasets.
Hardware Specification No The paper specifies components like 'Vi T-G version of vision transformer from EVA-CLIP' and 'Vicuna (13B) as the text generator' but does not provide specific hardware details such as GPU models, CPU types, or memory used for experiments.
Software Dependencies No The paper mentions 'Mini GPT-4', 'Vicuna (13B)', 'Med CLIP', and 'Adam W' as tools used, but it does not specify version numbers for these software components or other key dependencies like Python, PyTorch/TensorFlow, or CUDA.
Experiment Setup Yes The batch sizes for IU X-RAY and MIMIC-CXR are set to 12. The weights to balance I3 and C2FD loss in Eq. (2) are set to β1 = 1 and β2 = 1, respectively. In training, we only update parameters in the linear projection layer in the visual encoder and Vicuna through Adam W (Kingma and Ba 2015) with learning rate set to 1 10 6.