RadiologyNET

RadiologyNET is a family of convolutional neural network (CNN) foundation models pretrained on a large, in-house corpus of medical radiology images, intended as domain-specific alternatives to the ubiquitous ImageNet weights used for transfer learning in medical imaging. The work was developed primarily at the University of Rijeka (Faculty of Engineering and its Center for Artificial Intelligence and Cybersecurity) in Croatia, in collaboration with the Medical University of Graz in Austria, with imaging data sourced from the Clinical Hospital Centre Rijeka.

Most medical-imaging deep learning pipelines initialize from weights trained on natural photographs (ImageNet), despite the large domain gap between everyday images and grayscale radiographs, CT, and MR scans. RadiologyNET asks whether pretraining directly on radiology data yields better, more transferable representations. Because the underlying clinical archive was unlabeled, the authors first built an automated labeling pipeline (described in the 2024 BioData Mining paper) that data-mines three signals — image content, DICOM metadata, and narrative diagnosis text — and clusters their fused embeddings to annotate the dataset without manual labels.

The foundation models themselves and the transfer-learning evaluation were published in Scientific Reports in July 2025 under the title "Lessons learned from RadiologyNET foundation models for transfer learning in medical radiology," with pretrained weights released publicly for the community.

#Key Features

• Multimodal radiology pretraining: Models are pretrained on 1,902,414 DICOM-derived images spanning multiple anatomical regions and modalities (MR, CT, computed radiography, radiofluoroscopy, and angiography).
• Unsupervised label generation: Training labels were created automatically by clustering fused embeddings of image, DICOM-metadata, and free-text diagnosis features, avoiding costly manual annotation.
• Broad architecture coverage: Pretrained weights are provided for ten widely used CNN backbones, letting practitioners swap RadiologyNET initialization into existing pipelines.
• Drop-in transfer learning: A transfer_weights_to_model() utility loads the released weights across differing network topologies for fine-tuning on downstream tasks.
• Interpretability advantage: In the published study, radiologists judged RadiologyNET heatmaps to be the most dependable overall, with the best focus on pathological regions when present.

#Technical Details

RadiologyNET pretraining covers ResNet18, ResNet34, ResNet50, VGG16, EfficientNetB3, EfficientNetB4, InceptionV3, DenseNet121, MobileNetV3Small, and MobileNetV3Large, all implemented in PyTorch. The precursor dataset of roughly 1.34 million images was annotated into 50 clusters of visually similar images via multimodal feature fusion, with the full pretraining corpus reaching about 1.9 million images. The authors benchmarked RadiologyNET against ImageNet initialization and random initialization across several publicly available medical datasets, including classification, segmentation, and regression tasks. In most settings RadiologyNET and ImageNet performed almost identically when training was not data- or time-constrained, while RadiologyNET showed advantages under resource-limited conditions; multi-modality pretraining generally improved generalization.

#Applications

RadiologyNET is aimed at researchers and developers building medical-imaging models who want a domain-matched starting point for fine-tuning rather than natural-image weights. The released backbones can be fine-tuned for radiological classification, segmentation, and regression tasks, and are particularly attractive when labeled data or compute budgets are limited — a common situation in clinical research. Because weights are provided for ten standard architectures, teams can adopt RadiologyNET with minimal changes to existing training code.

#Impact

RadiologyNET contributes empirical evidence to an active debate about whether radiology-specific pretraining outperforms ImageNet transfer learning, alongside efforts such as RadImageNet. Its candid finding — that domain pretraining matches rather than uniformly beats ImageNet except in constrained or interpretability-sensitive settings — is a useful reference for the field, and the publicly released multi-backbone weights and unsupervised labeling methodology lower the barrier for groups exploring domain-specific foundation models. The work's main limitations are that the pretraining data originate from a single clinical center and that gains over ImageNet are situational rather than universal.