MOSAIC: A Multilingual, Taxonomy-Agnostic, and Computationally Efficient Approach for Radiological Report Classification

TL;DR

MOSAIC addresses the annotation bottleneck in radiology by enabling multilingual, taxonomy-agnostic classification of radiology reports with compact language models. It combines prompt-based zero-/few-shot capabilities with lightweight finetuning (LoRA) on consumer GPUs, demonstrating robust performance across English, Spanish, French, and Danish datasets and across chest X-ray and MRI modalities. The approach shows strong data efficiency, achieving expert-level macro-F1 on several chest X-ray tasks with relatively small annotated sets and leveraging data augmentation and cross-language transfer to further improve results. Open-source releases of 4B and 12B variants, together with comprehensive evaluation on multilingual, multi-taxonomy datasets, position MOSAIC as a practical alternative to large proprietary LLMs for clinical deployment.

Abstract

Radiology reports contain rich clinical information that can be used to train imaging models without relying on costly manual annotation. However, existing approaches face critical limitations: rule-based methods struggle with linguistic variability, supervised models require large annotated datasets, and recent LLM-based systems depend on closed-source or resource-intensive models that are unsuitable for clinical use. Moreover, current solutions are largely restricted to English and single-modality, single-taxonomy datasets. We introduce MOSAIC, a multilingual, taxonomy-agnostic, and computationally efficient approach for radiological report classification. Built on a compact open-access language model (MedGemma-4B), MOSAIC supports both zero-/few-shot prompting and lightweight fine-tuning, enabling deployment on consumer-grade GPUs. We evaluate MOSAIC across seven datasets in English, Spanish, French, and Danish, spanning multiple imaging modalities and label taxonomies. The model achieves a mean macro F1 score of 88 across five chest X-ray datasets, approaching or exceeding expert-level performance, while requiring only 24 GB of GPU memory. With data augmentation, as few as 80 annotated samples are sufficient to reach a weighted F1 score of 82 on Danish reports, compared to 86 with the full 1600-sample training set. MOSAIC offers a practical alternative to large or proprietary LLMs in clinical settings. Code and models are open-source. We invite the community to evaluate and extend MOSAIC on new languages, taxonomies, and modalities.

Figures (9)

• Figure 1: Performance as $(+)F1$ score of MedGemma-4B, Llama-8B, and Gemma-12B fine-tuned on mp$_{E+S}$c, with detailed results on three key Chest X-ray pathologies: Cardiomegaly, Pneumothorax, and Pleural Effusion. The data is presented across a range of chest X-ray datasets, illustrating model-specific and dataset-specific performance. On these common findings, all tested models have similar results, also generalizing on out-of-distribution datasets.
• Figure 2: Evaluation of different pretraining and fine-tuning configurations of MOSAIC on the DanskCXR. Fine-tuning directly on d dataset outperforms zero-shot and task pretraining.
• Figure 3: Data ablation study on DanskCXR shows that using just 480 examples (30% of the full dataset) is sufficient to reach peak performance.
• Figure 4: Performance of MOSAIC on the DanskMRI dataset, measured as $(+)F1$ across three epilepsy-related abnormalities from MRI reports.
• Figure 5: Distribution of labels in the chest X-ray datasets, by mention type and dataset.