Fig. 3: Overall weakly supervised training pipeline for critical findings extraction and performance validation. | npj Digital Medicine

Fig. 3: Overall weakly supervised training pipeline for critical findings extraction and performance validation.

From: Weakly supervised language models for automated extraction of critical findings from radiology reports

Fig. 3

The complete pipeline is composed of the following modules: a Phase I : This module generates the weak labels using unsupervised instruction-tuned Mistral class of LLMs; b Phase II: This module demonstrates the weakly supervised fine-tuning of the Mistral LLMs using the weak labels generated in Phase I; c This is module shows the expansion of critical findings terms and keywords. This expanded list is used for the key-based critical finding extraction; d This final pipeline demonstrates the model performance validation using internal and external datasets.

Back to article page