A block diagram of the Phase 02 models is provided below. Two of the models use the similar fine-tuning methods, and as such, we provide a demo notebook for fine-tuning, providing extended comments on the code. Additionally, since we employ pretraining before fine-tuning, we include a demo notebook for pretraining. To view the actual source code (*.py files) used for training and classifying, please refer to the train and classify documentation.


In the table below, we provide a description of the different tasks and architectures of the models.

Model Name



Source Code Function

Lung Findings, Adrenal Findings, or No Findings

Identify if the report has (i) lung (ii) adrneal, or (iii) or no findings

MLM based on RoBERTa (🤗)


Lung Recommended Procedure (i.e., Chest CT or Ambiguous Follow- up Recommended)

Identify if a report with lung has findings recommends (i) a Chest CT follow-up or (ii) some other ambiguous procedure

MLM based on RoBERTa (🤗)


Comment Extraction

Identify the portion of the report that contains the findings text

Question-Answer based RoBERTa (simpletransformers via 🤗)