Detecting Lung Cancer with Artificial Intelligence (Part 3): Sonador Lung AI
Lung cancer claims more lives each year than any other cancer. About 75% of those who have it will die within five years of their diagnosis. For lung cancer patients, early detection offers a window of opportunity — a chance to begin treatment before the cancer spreads, significantly improving the odds of survival. Right now, only about 15% of lung cancer cases are caught in those precious early stages, but when they are, the five-year survival rate jumps to 56%. This is the promise of AI: the potential to find these nodules when they are still manageable, before they evolve into something far more dangerous.
Tumors often go undetected in these critical early stages because traditional screening and diagnostic methods have limitations. X-rays, while widely available, struggle to capture small nodules with precision, and while CT scans provide clearer images, they are not used as widely for routine screening. Even with imaging, the visual cues can be so subtle that even experienced radiologists may miss them. AI aims to overcome these obstacles by enhancing detection and providing clinicians with a sharper, more reliable lens through which to identify early signs of lung cancer.
Sonador Lung AI is an initiative to create a solution for the early detection of lung cancer. It seeks to enable identification of lung nodules from both x-ray and low-dose CT, allow for the measurement of size and other quantitative characteristics to guide assessment/treatment, and provide seamless integration with clinical systems to report findings and manage care.
Complementary Imaging: X-ray and CT for Comprehensive Surveillance
No single modality is able to detect early lung cancer nodules by itself. For that reason, Sonador Lung AI will integrate X-ray and CT scans into a single solution. X-rays, widely available in hospitals and clinics worldwide, provide an initial, low-cost option for identifying potential nodules, making them ideal for frequent, broad surveillance. CT scans, on the other hand, offer higher resolution and more detailed imagery, enabling a closer look at suspicious findings to improve diagnostic accuracy.
By leveraging both imaging modalities, Sonador Lung AI’s multi-modality models will deliver more accurate and reliable detection and support comprehensive patient care pathways. We also wish to explore ways in which multi-channel and "stacked" models can help improve detection.
Project Goals
Early identification. Measure and inform. Seamless management.
Early Identification
Provide AI models able to detect and classify the presence of lung nodules on both x-rays and CT scans. X-rays because they are everywhere, and CT scans because they improve precision.
Measure and Inform
Quantify nodule size, shape, and growth rate to provide insight for risk assessment and to inform decisions on follow-up and intervention. Align with established protocols like the Fleischner Society guidelines and LUNG-RADS v2022.
Seamless Management
Integrate with PACS, RIS, EHR, and reporting systems to help track, report, and improve the delivery of care.
Solution Blueprint
Sonador Lung AI leverages the robust architecture of the Sonador platform, combining the input/output management of Sonador IO and the AI capabilities of Sonador AI to create an integrated platform for detecting, measuring, and managing early lung cancer indicators. This approach enables the storage and handling of medical imaging data in a secure fashion; the creation of image processing pipelines for data preparation; and the efficient training, application, and storage of AI models for use in clinical environments.
What Sonador Lung AI Needs to Deliver
To enable an end to end care solution, Sonador Lung AI needs to facilitate image acquisition and data import to PACS, provide a way to prepare images for assessment, apply models to assess cancer risk, and report findings.
1. Acquire Imaging
The system will need to gather thoracic imaging from hospital and clinic scanners to start the detection process and will need to support both chest X-rays and CT scans.
2. Import Medical Images to PACS
Once the imaging is collected, it should flow securely into the Sonador Imaging Platform for storage and preparation. The platform will use Orthanc as a PACS (Picture Archiving and Communication System) to ensure that the imaging studes are accessible for analysis.
3. Prepare Images for Assessment
The system must then optimize these images for AI-driven evaluation coordinated by an "AI Orchestrator" to prepare the data—whether X-ray, CT, or both—to ensure it is clean, formatted, normalized and ready for analysis by the AI model.
4. Assess Lung Cancer Risk
Sonador Lung AI will evaluate the imaging study's potential for malignancy as part of its core detection capabilities.
- Determine Malignancy Risk. By examining nodule characteristics such as size and growth patterns, the system needs to provide a reliable malignancy likelihood score that radiologists can trust. This score should align with established guidelines like the Fleischner Society and Lung-RADS v2022 for consistency and clinical relevance.
- Quantify Nodule Characteristics. For accurate assessment, the system should quantify each nodule’s defining features, giving clinicians the data needed to guide next steps in treatment.
5. Report Findings
Finally, Sonador Lung AI will provide access to the AI scores and results.
- Automate Report Creation.The system should generate structured, DICOM-compatible reports automatically, reducing the burden on clinicians and accelerating diagnosis. DICOM-SR will be used for encoding results to enable compatibility with PACS and other systems.
- Communicate Results to PACS, RIS, EHR. It will integrate seamlessly with PACS, RIS, and EHR systems to push results directly to the platforms radiologists already use, making the data available without requiring changes to their workflow.
Integrating AI Through Standards
In alignment with the 2024 RSA blueprint for AI in clinical settings, Sonador Lung AI adheres to established protocols and frameworks to integrate into healthcare workflows. DICOM, the standard for medical image exchange, is used to ensure consistent, secure transfer of imaging data, allowing AI-generated findings to be embedded alongside the original imaging series. Integrating the Healthcare Enterprise (IHE) profiles are used to model data flow and facilitate the receiving of imaging orders, acquisition of images from scanning modalities, initiate and coordinate the AI-driven processing for model inference, and push the results back to clinical systems for review. OpenEHR and FHIR are used to help facilitate communication with EHR and RIS.
Sonador Lung AI Platform
At the heart of Sonador Lung AI are powerful, interconnected components that work seamlessly together to deliver precise, accessible insights for early lung cancer detection.
Orthanc: Medical Database and Imaging Archive
Stores and organizes vast amounts of DICOM medical imaging data, ensuring high availability and easy access for model training and clinical application. The database also receives model results from the AI Orchestrator in DICOM-SR format to combine with source imaging study.
Sonador: Imaging Integration
Facilitates secure access to the platform, managing user roles, permissions, and integration across hospital systems to ensure privacy and compliance.
Kafka: Real-time Messaging
Provide real-time messaging to allow for data within the detection pipeline to be routed to other systems in the workflow and clinical environment.
Aranei: Aggregation, Enrichment, and Clinical Integration
Receive DICOM formatted AI results from Orthanc and translate/enrich them to OpenEHR/FHIR to facilitate communication with EHR, RIS, and critical findings systems.
Airflow: AI Orchestrator
Coordinates complex AI workflows, automating tasks like data processing, model training, and inference to maintain efficiency and consistency. Integrates with both Orthanc, Sonador, and MLflow and provides the runtime environment for MONAI and PyTorch.
MONAI: Image Preparation Pipelines
Provides advanced image processing tools to prepare CT and X-ray scans, enhancing the quality of input data used for model training and inference.
PyTorch: Deep Learning Framework
Serves as the backbone for developing deep learning models, enabling the creation and fine-tuning of sophisticated AI algorithms for nodule detection and measurement.
MLflow: Model Registry
Manages model versions and deployment, allowing for efficient tracking, storage, and retrieval to ensure consistent inference in clinical use.
Training AI for Tomorrow's Care
Sonador Lung AI uses an open ended architecture that allows for it to improve over time. Models can be re-trained from real-world imaging data stored in Orthanc (which can be verified by radiologists to ensure accuracy). As new model architectures or detection technologies become available, they can be incorporated into the detection model to improve accuracy.
The figure below illustrates how a new model instance can be created. Imaging scans (in DICOM format) along with diagnostic labels that describe what the image should ideally reveal are retrieved from Orthanc. MONAI pipelines are used to combine the imaging and label data, creating a dataset that AI can use to learn patterns and improve its detection capabilities.
To train the model, this dataset is divided into three groups: training data, which the model uses to learn; testing data, which helps monitor its accuracy; and validation data, which ensures the model is reliable and unbiased. Through a process called a "training loop," the model repeatedly analyzes the data until it can reliably identify nodules or other areas of concern. Once the model reaches a high level of accuracy, it is saved and stored in a model registry, ready for use in future scans.
Because model instances are stored in MLflow along with information about their source data and training parameters, it is possible to train multiple models (with slightly different parameters) and assess which perform best. Likewise, should model accuracy decrease over time, models can be compared to previous versions to track data drift.
Sonador Lung AI follows Sonador AI conventions for tracking data lineage, which aids in capturing the details of source data and training parameters in MLflow.
Applying Sonador Lung AI Models
Once a model has been generated, it can applied to new imaging studies. The figure below shows how predictions are generated for new studies. Imaging data is retrieved from Orthanc in DICOM format. Sonador Lung AI prepares images for analysis by executing a processing pipeline which will convert them to a specific format; combine the data into "channels" for the model to interpret; normalize pixel intensities and remove artifacts; and ensure that a consistent "coordinate space" is used.
Once prepared, the imaging data is passed to a PyTorch model which will use AI to identify signs of lung cancer. The model will also quantify tumor characteristics using Fleischner society and Lung-RADS v2022 guidelines. AI predicted scores and tumor characteristics are encoded using DICOM-SR, with UID links to the images allowing Sonador to route the results to clinical systems, such as PACS, RIS, and EHR.
Proof of Concept and Future Development
Oak-Tree, along with members of the Sonador community completed a successful proof of concept of the architecture and data flows using CT scans from the "Lung Image Database Consortium (LIDC) and Image Database Resource Initiative (IDRI)" (commonly referred to as the LIDC-IDRI dataset) earlier this year.
Using a convolutional neural network, we were able to achieve accuracy similar to that reported in the clinical literature. We also showed that Sonador Lung AI was able to process patients concurrently and that candidate models could be trained and used for inference in a cost effective manner (using cloud and on-premise deployments of Sonador). A detailed report which will include the model architecture, results, and source code will be published separately.
Next steps include:
- Modifying the architecture of the existing model to support multiple input channels (similar to the approach reported by Ardila et al), allowing it to accommodate prior scans and x-ray input.
- Assessing transformer based architectures against the multi-channel neural network model.
We hope to deliver source code, model weights, and deployment documentation to the Sonador community soon to allow for the candidate models to be more broadly reviewed and assessed.
Identify, Measure, and Manage
Sonador Lung AI aims to reshape the landscape of early lung cancer detection by merging cutting-edge AI with the familiar rhythms of clinical practice. It will be a transparent, open source, integrated, end-to-end solution to facilitate imaging acquisition and preparation, cancer risk assessment, and streamlined reporting. The potential impact of catching lung cancer in its earliest stages is profound, providing significant benefits to both patients and health systems.
Comments
Loading
No results found