---
title: 'Neu-RadBERT: Radiology NLP for Brain Reports'
url: https://www.emergentmind.com/topics/neu-radbert
type: topic
---

# Neu-RadBERT: Radiology NLP for Brain Reports

Neu-RadBERT is a BERT-based model developed to extract diagnoses from free-text radiology reports of brain imaging performed in patients with acute respiratory failure undergoing invasive mechanical ventilation. In the reported formulation, the model is fine-tuned for multi-label classification of unstructured brain CT and MRI reports, with three training strategies beyond a baseline RadBERT comparator: supervised fine-tuning, additional masked language modeling pretraining, and oversampling to address data skewness. The study positions Neu-RadBERT as a target-domain radiology NLP system for converting narrative reports into structured neurological labels, and reports substantial gains over baseline RadBERT and Llama-2-13B on the specified task [2510.06232].

## 1. Clinical task and research setting

The immediate motivation for Neu-RadBERT is the need to automatically extract structured diagnostic information from free-text brain radiology reports in critically ill patients, particularly those with acute respiratory failure undergoing invasive mechanical ventilation. Manual extraction and rule-based systems are described as laborious, time-consuming, and not scalable. The study further notes that transformer-based models show promise for automating classification from free text, but that general pre-trained models or general LLMs lack the domain specificity and performance needed for fine-grained diagnostic tasks, especially when labels are not mutually exclusive and class imbalances are present [2510.06232].

Neu-RadBERT sits within a broader radiology-language-model trajectory in which domain adaptation has repeatedly been associated with improved downstream performance. In automatic CT protocol assignment, a domain-specific $BERT_{rad}$ achieved a macro-averaged F1 of 0.63 versus 0.61 for $BERT_{base}$, and knowledge distillation improved the score to 0.66, with the gains concentrated in minority classes [2009.00694]. In brain MRI report classification, fine-tuned BERT, BioBERT, ClinicalBERT, and RadBERT achieved F1-Score \(>97\%\), and RadBERT slightly led, particularly on adult out-of-distribution data [2503.12143]. This suggests that Neu-RadBERT extends an established radiology-NLP pattern: task performance improves when the encoder is adapted to the report domain and to the label distribution of the target problem.

## 2. Corpus construction and label space

The Neu-RadBERT study used the MIMIC-IV and MIMIC-IV-Note databases from Beth Israel Deaconess Medical Center, covering 2008–2019. The source pool consisted of all brain CT and MRI reports for patients with acute respiratory failure requiring invasive ventilation and without pre-existing neurological injury on ICU admission. From 219,532 eligible reports, 1,200 reports were randomly sampled for labelling [2510.06232].

The annotation schema comprised 11 non-mutually-exclusive labels: Chronic Brain Injury, Acute Brain Injury, Ischemic Stroke, Subdural Hemorrhage, Subarachnoid Hemorrhage, Intraparenchymal Hemorrhage, Intraventricular Hemorrhage, Anoxic-hypoxic Brain Injury, Brain Edema, Microbleeds, and Intracranial Hypertension. Two experts independently labelled each report, and discrepancies were resolved by a senior author. The initial split allocated 1,000 reports into train \((n=800)\) and validation/test \((n=200)\), while an additional 200 reports were labelled for oversampling rare classes. For oversampling, keyword-based free-text searching was used to identify rare conditions such as microbleeds, with manual confirmation before inclusion [2510.06232].

The study also reports a limited preprocessing step specific to Llama-2 fine-tuning: removal of newline characters. No comparable task-specific normalization procedure is described for the BERT-based pipeline. The overall design therefore emphasizes label curation and domain adaptation rather than extensive handcrafted preprocessing.

## 3. Model lineage and training variants

Neu-RadBERT is defined relative to a baseline RadBERT model. RadBERT is described as a standard BERT-based transformer pretrained on millions of VA radiology reports and initially used out-of-the-box with zero additional training on the target data. Neu-RadBERT preserves the same basic configuration—12 Transformer layers, 12 attention heads, and hidden size 768—and modifies training rather than architectural depth or width [2510.06232].

The reported variants are summarized below.

| Variant | Procedure | Data change |
|---|---|---|
| A0 | Baseline RadBERT | Zero additional training |
| A1 | Fine-tuned Neu-RadBERT | Supervised fine-tuning on 800 reports |
| A2 | Neu-RadBERT with MLM pretraining | Unsupervised MLM on \(\sim 2{,}000\) unlabelled reports, then supervised fine-tuning |
| A3 | Neu-RadBERT with MLM pretraining and oversampling | Added 200 more labelled reports for rare classes |

Fine-tuned Neu-RadBERT used standard supervised cross-entropy loss for binary classification per label and was trained for 10 epochs. The A2 variant introduced target-domain masked language modeling, in which random word masking was used and the model predicted masked words so as to adapt more closely to the language of the target corpus. The A3 variant addressed class imbalance by identifying underrepresented labels and increasing their representation through additional manually confirmed examples, specifically improving coverage for microbleeds, subdural hemorrhage, anoxic injury, and intracranial hypertension. The oversampled set used an 80:20 train:validation split within the added data [2510.06232].

Within the broader radiology NLP literature, these choices are consistent with earlier findings that local or task-specific pretraining can be beneficial, and that imbalance-aware training materially affects minority-label performance. In CT protocol assignment, knowledge distillation combined with augmentation improved macro-F1 from 0.63 for $BERT_{rad}$ to 0.66 for BAN2/BAN3, while naïve undersampling reduced performance to 0.39 [2009.00694]. A plausible implication is that Neu-RadBERT’s design combines two previously productive strategies—domain adaptation and rare-class handling—within a brain-report extraction setting.

## 4. Reported performance on brain abnormality classification

Evaluation was reported on an external validation set of \(n=200\) reports. The baseline RadBERT condition underperformed substantially on several neurologic findings, whereas Neu-RadBERT improved accuracy across nearly all listed abnormalities after supervised fine-tuning, masked language modeling, and oversampling. For Chronic Brain Injury, accuracy increased from 84.1% in the RadBERT baseline to 88.7% with A1, 90.3% with A2, and 89.2% with A3. For Acute Brain Injury, the corresponding values were 45.1%, 99.0%, 99.5%, and 98.0%. For Ischemic stroke they were 54.9%, 95.9%, 96.4%, and 96.9%; for Subdural hemorrhage, 59.5%, 97.4%, 99.0%, and 99.5%; for Subarachnoid hemorrhage, 54.4%, 98.0%, 97.4%, and 99.0%; for Intraparenchymal hemorrhage, 56.4%, 94.4%, 96.4%, and 97.4%; for Intraventricular hemorrhage, 54.9%, 95.4%, 97.4%, and 97.4%; for Anoxic brain injury, 57.4%, 97.4%, 98.5%, and 98.5%; for Brain edema, 58.5%, 97.4%, 97.9%, and 97.4%; and for Intracranial hypertension, 57.2%, 96.4%, 97.4%, and 98.0%. Microbleeds were reported only for A3, at 98.0% [2510.06232].

The study also provides example A3 operating metrics beyond accuracy. For Acute Brain Injury, it reports Accuracy 99.17%, F1-score 0.9917, Precision 0.9916, and Recall 0.9917. For Subdural hemorrhage, it reports Accuracy 93.75%, F1-score 0.9073, Precision 0.8789, and Recall 0.9375 [2510.06232].

These results are presented as evidence that target-domain pretraining and oversampling alter performance most strongly where the baseline model is weakest, namely on highly specific neurological abnormalities and rare findings. The paper explicitly states that oversampling rare categories is critical for robust recognition of infrequent but clinically crucial findings such as microbleeds.

## 5. Comparison with general-purpose LLMs

Neu-RadBERT was compared with Llama-2-13B, treated as an autoregressive LLM baseline for the classification task. On acute-versus-not-acute binary classification, Llama-2-13B achieved 26.7% accuracy in zero-shot mode, 34.2% in in-context mode, 67.5% in fine-tuned zero-shot mode, and 66.7% in fine-tuned in-context mode. By contrast, Neu-RadBERT with oversampling was reported to achieve up to 98.0% accuracy for acute brain injuries, while the broader Neu-RadBERT conclusion emphasized significant improvements in diagnostic accuracy compared to baseline RadBERT for brain abnormalities [2510.06232].

The study attributes Llama-2-13B’s relative underperformance to several factors: both in-context and fine-tuned models were described as overconfident and/or collapsed to a single class in many cases, and its embeddings and vocabulary were described as not well-adapted to the radiology domain. At the same time, the paper does not frame this as a universal limit of autoregressive LLMs; it explicitly notes that larger models or further fine-tuning could improve performance [2510.06232].

This comparison intersects with a recurrent issue in clinical NLP: model scale does not by itself guarantee superior task performance. In pediatric brain MRI report classification, Gemini 1.5-Pro showed promising categorization performance on a difficult error subset, especially with stepwise clinical reasoning, but the BERT-family systems remained the primary high-throughput classifiers for the full dataset [2503.12143]. A common misconception is therefore that a larger general LLM necessarily dominates a domain-adapted encoder on radiology classification; the Neu-RadBERT results argue against that assumption for this specific diagnostic extraction problem.

## 6. Relation to adjacent radiology-language models

Neu-RadBERT is best understood as part of a family of radiology-adapted BERT systems rather than as an isolated architecture. One antecedent is $BERT_{rad}$ for automatic assignment of radiology examination protocols, where the input combined structured and unstructured fields in a semi-structured sequence and further in-domain pretraining improved over general-domain BERT. That study also showed that imbalance mitigation through augmentation and Born-Again Neural Network distillation improved minority-class behavior without improving the weighted micro-average, underscoring the importance of reporting class-sensitive metrics under skewed label distributions [2009.00694].

A second adjacent line is automated classification of brain MRI reports as normal or abnormal. There, fine-tuned RadBERT and related BERT variants achieved robust performance under unbalanced training, with full report text outperforming impression-only inputs: for RadBERT, Full Report F1 was \(97.06 \pm 0.58\) versus Impression-only F1 of \(91.52 \pm 0.39\). The same study found near-perfect agreement between language-model-derived and human-derived brain growth charts, with correlation \(r = 0.99\) and \(p < 2.2e^{-16}\) [2503.12143].

Taken together, these antecedents place Neu-RadBERT in a convergent methodological space. This suggests that its reported gains are not solely attributable to the choice of BERT backbone, but to a composite of radiology-specific pretraining, careful label design, and explicit management of class imbalance.

## 7. Applications, limitations, and availability

The principal application proposed for Neu-RadBERT is the transformation of large-scale free-text radiology corpora into structured data labels for research and downstream machine learning. The paper identifies potential uses in causal inference, epidemiology, and as weak labels for computer vision. It also argues that lightweight BERT-based models are feasible to deploy on-premises, which may address privacy and HIPAA concerns that hinder some LLM-based applications [2510.06232].

The study is also explicit about its limitations. External validation on other institutional datasets is needed for broader generalizability. It calls for exploration of other, possibly medical and non-autoregressive, LLMs, along with more parameter-efficient fine-tuning strategies. It further notes the possibility of using automatically labelled data to train computer vision algorithms for imaging-based diagnosis [2510.06232].

Neu-RadBERT’s reported public release is through the Hugging Face repository at `https://huggingface.co/datasets/manisggn/Neu-Radbert/tree/main`. In the surrounding literature, similar report classifiers have already been connected to secondary pipelines such as automated brain growth chart generation, where language-model classification can serve as a gatekeeping stage for quantitative imaging analysis [2503.12143]. A plausible implication is that Neu-RadBERT may be most consequential not as a standalone classifier, but as an infrastructural component in larger neurocritical care and neuroimaging data workflows.

Source: https://www.emergentmind.com/topics/neu-radbert