Chest X-ray analysis,
open to all.

What is Open Opinion?

Open Opinion is an independent research project developing an open-source AI system for chest X-ray analysis. The system is designed to identify thoracic conditions from frontal chest X-ray images, providing a probability score for each of 17 conditions simultaneously along with a heatmap showing which regions of the image influenced the prediction.

The project is entirely self-funded and uses only publicly available datasets. All training code, methodology and results are documented here transparently.

The Training Process

The system was built using a technique called transfer learning — starting from a ResNet-50 neural network pre-trained on 14 million photographs, then fine-tuning it to recognise chest pathology. The full process involved:

1. Dataset assembly — Sourcing and combining seven publicly available chest X-ray datasets, totalling over 156,000 images across 17 conditions.
2. Data preprocessing — Resizing all images to 224×224 pixels, applying ImageNet normalisation, and writing custom parsers to handle DICOM, NIfTI, PNG and CSV annotation formats.
3. Class balancing — Applying inverse-frequency class weighting in the loss function to ensure rare conditions like Hernia and Fibrosis are not ignored during training.
4. Augmentation — Applying conservative augmentations during training including small rotations, brightness/contrast variation and Gaussian noise to improve generalisation.
5. Training — Running multiple training iterations on NVIDIA L40S GPU instances, progressively adding datasets and monitoring per-condition F1 scores after each run.
6. Transfer learning between datasets — Using each trained model as the starting point for the next training run, so no learning is lost between dataset additions.
7. Evaluation — Holding out 10% of data as an unseen test set, evaluating accuracy, macro F1 and per-condition F1 scores.

Limitations

Open Opinion is a research prototype. It is not approved for clinical use and must not be used for diagnostic purposes. The system performs well on common, well-represented conditions but struggles with rarer findings where limited training data is available. Performance will continue to improve as more annotated data is incorporated.

Conditions such as Consolidation and Atelectasis remain challenging because they are visually similar to other conditions — a difficulty shared by human radiologists. The confusion between Pneumonia and Consolidation in particular is a known challenge in the radiology literature.

What's next

Pending approval from PhysioNet, the system will be trained on MIMIC-CXR (227,000 images) and VinDr-CXR (18,000 expert-annotated images with bounding boxes). These datasets are expected to significantly improve accuracy, particularly for conditions currently limited by data volume.

A radiologist review pipeline is also under development, allowing expert corrections to feed back into model retraining over time.