Awesome

Improving the Factual Correctness of Radiology Report Generation with Semantic Graph-Based Rewards

Rebuttal

Please find the updated manuscript in "updated_manuscript.pdf" <br/> Also, you can evaluate our best result on mimic in "eval_mimic.ipynb" with output: <br/>

{
    "ROUGEL": 0.25957182933436834,
    "radgraph_simple": 0.3707634326584056,
    "radgraph_partial": 0.3473400863906501,
    "radgraph_complete": 0.23481543928467108,
    "radentitymatchexact": 0.42533178285691237,
    "radentitynli": 0.4329851069563082,
    "chexbert-5_micro avg_f1-score": 0.6215361255219537,
    "chexbert-all_micro avg_f1-score": 0.5600310338139264
}

First submission

RadGraph.ipynb will evaluate the output of a system against the reference report using the RadGraph metric (RG_ER and RG_\hat{ER}). It also prints the sets \hat{V}_\hat{y} and \hat{V}_y (similarly to Appendix B) <br/>

To not violate the double blind review process, we do not release the training code for now. RadGraph can be easily integrated in any RRG framework given the code in RadGraph.ipynb

erratum: Appendix C 'study' :

Reader 1: 48.5 ± 66.2 
Reader 2: 89.1 ± 46.1

to

Reader 1: 0.485 ±0.662 
Reader 2: 0.891 ± 0.461