language:
- en
metrics:
- rouge
library_name: transformers
pipeline_tag: summarization
tags:
- medical
- radiology
- biobart
- impression
- findings
- biomedical
- bart
datasets:
- MIMIC-III
widget:
- >-
post contrast axial sequence shows enhancing large neoplasm left parietal
convexity causing significant amount edema mass effect study somewhat
limited due patient motion similar enhancing lesion present inferior aspect
right cerebellar hemisphere right temporal encephalomalacia noted mra brain
shows patent flow anterior posterior circulation evidence aneurysm vascular
malformation
- >-
seen hypodensity involving right parietal temporal lobes right cerebellar
hemisphere effacement sulci mild mass effect lateral ventricle hemorrhage
new region territorial infarction basal cisterns patent mucosal thickening
fluid within paranasal sinuses aerosolized secretions likely related
intubation mastoid air cells middle ear cavities clear
- >-
heart size normal mediastinal hilar contours unchanged widening superior
mediastinum likely due combination mediastinal lipomatosis prominent thyroid
findings unchanged compared prior ct aortic knob mildly calcified pulmonary
vascularity engorged patchy linear opacities lung bases likely reflect
atelectasis focal consolidation pleural effusion present multiple old
rightsided rib fractures
inference:
parameters:
max_length: 350
Model Card for Model ID
The model can be used to generate accurate and informative radiology impressions, which can help to improve communication between radiologists and other healthcare providers.
Model Details
Model Description
This model is a BioBart-based sequence-to-sequence model that was trained on the custom dataset to summarize radiology findings into impressions. In training i use 70000 radiology reports to train the model to summarize radiology findings into impressions.
- Developed by: [Engr. Hamza Iqbal Malik (UET TAXILA)]
- Shared by : [Engr. Hamza Iqbal Malik (UET TAXILA)]
- Model type: [Medical Text Summarization Model]
- Language(s) (NLP): [English]
- Finetuned from model: [GanjinZero/biobart-v2-base]
Model Sources
- Repository: [GanjinZero/biobart-v2-base]
- Paper: [BioBART: Pretraining and Evaluation of A Biomedical Generative Language Model]
- Demo: [hamzamalik11/radiology_summarizer]
Uses
The model is intended to be used for generating radiology impressions. The model should not be used for any other purpose.
Direct Use
The model can be directly used for generating impressions from radiology reports. Users can input the findings of a radiology report, and the model will generate a summarized impression based on that information.
Out-of-Scope Use
The model should not be used for any purpose other than generating impressions from radiology reports. It is not suitable for tasks outside of radiology report summarization.
Recommendations
Users should be aware of the limitations and potential biases of the model when using the generated impressions for clinical decision-making. Further information is needed to provide specific recommendations.
How to Get Started with the Model
Use the code below to get started with the model. from transformers import AutoTokenizer, AutoModelForMaskedLM, AutoModelForSeq2SeqLM from transformers import DataCollatorForSeq2Seq
model_checkpoint = "attach your trained model here"
model = AutoModelForSeq2SeqLM.from_pretrained(model_checkpoint) tokenizer = AutoTokenizer.from_pretrained(model_checkpoint) from transformers import SummarizationPipeline
summarizer = SummarizationPipeline(model=model, tokenizer=tokenizer)
output= summarizer("heart size normal mediastinal hilar contours remain stable small right pneumothorax remains unchanged surgical lung staples overlying left upper lobe seen linear pattern consistent prior upper lobe resection soft tissue osseous structures appear unremarkable nasogastric endotracheal tubes remain satisfactory position atelectatic changes right lower lung field remain unchanged prior study")
Training Details
Training Data
The training data was a custom dataset of 70,000 radiology reports.The data was cleaned to remove any personal or confidential information. The data was also tokenized and normalized. The training data was split into a training set and a validation set. The training set consisted of 63,000 radiology reports, and the validation set consisted of 7,000 radiology reports.
Training Procedure
The model was trained using the Hugging Face Transformers library: https://huggingface.co/transformers/. The model was trained using the AdamW optimizer with a learning rate of 5.6e-5. The model was trained for 10 epochs.
Training Hyperparameters
- Training regime: -[evaluation_strategy="epoch"], -[learning_rate=5.6e-5], -[per_device_train_batch_size=batch_size //4], -[per_device_eval_batch_size=batch_size //4,] -[weight_decay=0.01], -[save_total_limit=3], -[num_train_epochs=num_train_epochs //4], -[predict_with_generate=True //4], -[logging_steps=logging_steps], -[push_to_hub=False]
Evaluation
Testing Data, Factors & Metrics
Testing Data
The testing data consisted of 10,000 radiology reports.
Factors
The following factors were evaluated: [-ROUGE-1] [-ROUGE-2] [-ROUGE-L] [-ROUGELSUM]
Metrics
The following metrics were used to evaluate the model: -[ROUGE-1 score: 44.857] -[ROUGE-2 score: 29.015] -[ROUGE-L score: 42.032] -[ROUGELSUM score: 42.038]
Results
The model achieved a ROUGE-L score of 42.032 on the testing data. This indicates that the model was able to generate summaries that were very similar to human-written summaries.
Summary
The model was trained on a custom dataset of 70,000 radiology reports. The model was able to achieve a ROUGE-L score of 42.032 on the testing data. This indicates that the model is able to generate summaries that are very similar to human-written summaries.
Model Card Authors
-Name: Engr. Hamza Iqbal Malik
-LinkedIn: www.linkedin.com/in/hamza-iqbal-malik-42366a239
-GitHub: https://github.com/hamza4344
Model Card Contact
-Name: Engr. Hamza Iqbal Malik
-LinkedIn: www.linkedin.com/in/hamza-iqbal-malik-42366a239
-GitHub: https://github.com/hamza4344