Original model description:

language: - en metrics: - rouge-l tags: - medical - summarization - clinical - bart - Radiology - Radiology Reports 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

Radiology Report Summarization

This model summarizes radiology findings into accurate, informative impressions to improve radiologist-clinician communication.

Model Highlights

Model name: Radiology_Bart
Author: Muhammad Bilal
Model type: Sequence-to-sequence model
Library: PyTorch, Transformers
Language: English

Parent Model

Repository: GanjinZero/biobart-v2-base
Paper: BioBART: Pretraining and Evaluation of A Biomedical Generative Language Model

This model is a version of pretrained BioBart-v2-base model further finetuned on 70,000 radiology reports to generate radiology impressions. It produces concise, coherent summaries while preserving key findings.

Model Architecture

Radiology_Bart is built on the BioBart architecture, a sequence-to-sequence model which is pre-trained on biomedical-text-dataPubMed. The encoder-decoder structure allows it to compress radiology findings into impression statements.

Key components:

Encoder: Maps input text to contextualized vector representations
Decoder: Generates output text token-by-token
Attention: Aligns relevant encoder and decoder hidden states

Data

The model was trained on 70,000 deidentified radiology reports split into training (52,000), validation (8,000), and test (10,000) sets. The data covers diverse anatomical regions and imaging modalities (X-ray, CT, MRI).

Training

Optimization: AdamW
Batch size: 16
Learning rate: 5.6e-5
Epochs: 4

The model was trained to maximize the similarity between generated and reference impressions using ROUGE metrics.

Performance

Evaluation Metrics

ROUGE-1 score	ROUGE-2 score	ROUGE-L score	ROUGELSUM score
44.857	29.015	42.032	42.038

Demonstrating high overlap with human references.

Usage

from transformers import AutoTokenizer, AutoModelForSeq2SeqLM, Pipeline

# Sample findings 
findings = "There is a small lung nodule in the right upper lobe measuring 6 mm. The heart size is normal. No pleural effusion or pneumothorax."

# Load model & tokenizer
summarizer = pipeline("summarization", model="Mbilal755/Radiology_Bart")
tokenizer = AutoTokenizer.from_pretrained("Mbilal755/Radiology_Bart")

# Tokenize findings
inputs = tokenizer(findings, return_tensors="pt")

# Generate summary 
summary = summarizer(findings)[0]['summary_text']

# Print outputs
print(f"Findings: {findings}")
print(f"Summary: {summary}")

Limitations

This model is designed solely for radiology report summarization. It should not be used for clinical decision-making or other NLP tasks.

RichardErkhov
/

Mbilal755_-_Radiology_Bart-4bits