Disease probability-enhanced follow-up chest X-ray radiology report summary generation
A chest X-ray radiology report describes abnormal findings not only from X-ray obtained at a given examination, but also findings on disease progression or change in device placement with reference to the X-ray from previous examination. Majority of the efforts on automatic generation of radiology report pertain to reporting the former, but not the latter, type of findings.
To the best of the authors’ knowledge, there is only one work dedicated to generating summary of the latter findings, i.e., follow-up radiology report summary. In this study, we propose a transformer-based framework to tackle this task. Motivated by our observations on the significance of medical lexicon on the fidelity of report summary generation, we introduce two mechanisms to bestow clinical insight to our model, namely disease probability soft guidance and masked entity modeling loss. The former mechanism employs a pretrained abnormality classifier to guide the presence level of specific abnormalities, while the latter directs the model’s attention toward medical lexicon. Extensive experiments were conducted to demonstrate that the performance of our model exceeded the state-of-the-art.
Chest X-ray radiology report generation model can automatically report radiological findings on radiogram in an end-to-end manner, potentially providing a means to alleviate the workloads of radiologists. A major shortcoming of the majority of these models is that they are limited to reporting findings that are present in a single chest X-ray examination. However radiologist typically writes report based not only on a single (follow-up) X-ray but also on the X-ray from previous check-up to discern change in disease severity, and/or change in the placement of medical device. This type of radiology report is denoted as follow-up radiology report. In other words, majority of the current report generation models fail to report these disease progression-related findings, and are thus not suitable for the follow-up radiology report generation task.
Considering that automatic generation of a follow-up radiology report is a clinically relevant and important problem, this study aims to develop an end-to-end model to tackle a related and simplified generative problem, dubbed follow-up radiology report summary generation, whereby a textual summary of disease progression is generated based on a pair of chest X-rays obtained at follow-up and baseline examinations (see Fig. . There are two key challenges to the generation of follow-up report summary. The first is the diagnostic accuracy of the generated report summary. Take a ground truth follow-up report summary – The follow-up chest X-ray is missing the finding of pneumonia compared to the baseline X-ray – as an example. It is acceptable if the model mis-predicts the word “of”, but not acceptable when the model makes the wrong diagnosis of the abnormality – pneumonia. It is therefore of paramount importance to maintain the clinical accuracy of generated reports. The other issue is related to the suboptimal use of equal attention to each word in the generated report under the supervision of the commonly used masked language modeling (MLM) loss, e.g., As previously mentioned, the clinical accuracy of a generated report summary is more important than the fidelity of other generated words. Abnormalities-related entities should therefore deserve more attention than other words.
In this study, we therefore propose two mechanisms to partly address each of these two issues, whereby additional diagnostic information is provided, and model attention is reallocated in the hope of improving the prediction of medical lexicon. These mechanisms, respectively dubbed disease probability soft guidance and masked entity modeling loss, aim to bestow clinical insight on the follow-up report summary generation task. Disease probability soft guidance employs a pretrained abnormality classifier to provide complementary guidance on the presence level of abnormalities commonly seen on chest X-ray, whilst masked entity modeling loss directs the model’s attention toward abnormality-related words. In this study, our proposed framework was tested on two groups of follow-up findings, namely change in the presence or absence of abnormalities and change in the severity of abnormalities (see Table for the full list of follow-up findings being investigated). A preliminary version of this work has been reported in Our contributions are
Website: cognitivescientist.org
Comments
Post a Comment