Clinically Correct Report Generation from Chest X-Rays Using Templates

Evaluating Progress in Automatic Chest X-Ray Radiology Report Generation

“…To generate metric-oracles, any report generation model is theoretically feasible. There are three main categories: the first generates free text based on semantics extracted from input chest X-ray images 20,33,34 ; the second retrieves existing text that best matches input images from a report corpus 17,35 ; and the third selects curated templates corresponding to a predefined set of abnormalities 14,36 . We chose to use retrieval-based models to generate metric-oracles because retrieval from a training report corpus produces a controlled output space, instead of an unpredictable one produced by models that generate free text.…”

“…Namely, we can curate a set of medical conditions and obtain radiologist annotations for each condition over a training set of reports. Then, we can train a classifier that outputs the likelihood of having each condition given an X-ray image, and proceed to select the corresponding report templates for conditions with high likelihood 14 . Some more nuanced approaches paraphrase the curated templates after selection 36 .…”

“…Enabled by the rapidly evolving imaging and computer vision technologies, AI also has made formidable progress on image interpretation tasks, including chest X-ray interpretation. However, the application of AI to image interpretation tasks has often been limited to the identification of a handful of individual pathologies [12][13][14] , representing an oversimplification of the image interpretation task. In contrast, the generation of complete narrative radiology reports [15][16][17][18][19][20] moves past that simplification and matches up to how radiologists communicate diagnostic information: the narrative report allows for highly diverse and nuanced findings, including association of findings with anatomic location, and expressions of uncertainty.…”

Few-shot Structured Radiology Report Generation Using Natural Language Prompts

“…However, only a few works focus their research on structured report generation. Pino et al [25] proposed a model for structured reporting generation where abnormalities are first classified, and the correct report templates are automatically selected. To improve the structured reporting results, additional object detectors were proposed as an additional step [3].…”

“…With the rise of powerful Natural Language Processing (NLP) models like transformers [27] and the availability of large chest radiograph datasets, many methods have been proposed for the generation of free-text radiology reports. However, these generated reports suffer from the same drawbacks as humanwritten free-text reports: the terminology is not standardized, and therefore it is difficult to evaluate whether a report is clinically accurate or merely performs well on common NLP metrics which do not measure the expressiveness of a report in all its clinical details [24]. To this end, we propose a structured report generation method that requires less annotated data by leveraging pre-training of a contrastive language-image model on chest radiograph and report pairs.…”