Novel ML for Pulmonary Embolism Detection

Pulmonary embolism (PE) is a critical medical condition caused by a blockage in one of the pulmonary arteries in the lungs, usually due to blood clots. It is responsible for significant morbidity and mortality worldwide, particularly because symptoms can often be non-specific, making diagnosis challenging. Timely and accurate detection is crucial, as untreated PE can lead to life-threatening complications such as heart failure or sudden death.

CT pulmonary angiography (CTPA) is the gold standard imaging method for diagnosing PE, as it provides a detailed visualization of the pulmonary vasculature. However, current machine learning (ML) models designed to assist in PE detection are heavily reliant on detailed, manual slice-level annotations from radiologists. This process is not only labor-intensive and costly but also vulnerable to human error, which can result in inconsistent labeling. Additionally, clinical data can vary significantly across institutions, making it difficult to generalize the performance of existing models across different healthcare settings.

TECHNOLOGY

The semi-weakly supervised learning model is designed to overcome these challenges by reducing the need for extensive manual annotations while maintaining high diagnostic accuracy. Unlike traditional ML models, which often require granular slice-level labels, this model employs a combination of exam-level labels and a limited number of slice-level annotations to train effectively.

The model utilizes a one-stage, end-to-end architecture that integrates two core components:

• 2D Convolutional Neural Network (CNN): This is used for feature extraction from CT images, capturing important visual information about the pulmonary structures.
• Recurrent Neural Network (RNN): This component captures temporal relationships between imaging slices, helping the model understand sequential information, which is crucial for identifying PE that may affect different regions of the lungs over multiple slices.

Additionally, the model incorporates Grad-CAM (Gradient-weighted Class Activation Mapping) tools to provide visual explanations of its predictions. Grad-CAM helps clinicians better understand the regions of the images that the model focuses on when making decisions, particularly in detecting emboli in both central and peripheral areas of the pulmonary arteries.

BENEFITS

1. Reduced Annotation Burden: Traditional ML models require intensive manual labeling of each slice, but this model requires fewer detailed annotations, making it much more efficient and cost-effective to train.
2. Improved Generalizability: The model’s ability to adapt to variations in clinical data across different healthcare sites ensures more consistent performance, regardless of data source or location.
3. Enhanced Diagnostic Accuracy: By leveraging both exam-level and slice-level labels, the model maintains high accuracy, even with reduced data annotation, ensuring reliable detection of PE.
4. Scalability: This technology has the potential to be easily scaled and implemented in different healthcare environments due to its reduced reliance on manual data labeling and its robust performance across diverse datasets.
5. Increased Clinical Trust: The integration of Grad-CAM tools provides clinicians with visual explanations of the model’s decision-making process, improving transparency and trust in AI-driven diagnoses.
6. Lower Costs: Reducing the need for manual annotations directly cuts down costs associated with radiologist labor, making the technology more economically viable for widespread adoption.

APPLICATIONS

1. Pulmonary Embolism Detection: The primary application of this model is in detecting PE in CT pulmonary angiography scans, where timely and accurate detection can significantly improve patient outcomes.
2. General Diagnostic Imaging: While the current focus is on PE detection, the model’s architecture could be adapted for other diagnostic imaging tasks that rely on time-series or sequential data, such as detecting cancers, strokes, or other vascular conditions in various imaging modalities (CT, MRI).
3. Clinical Settings: Hospitals, diagnostic imaging centers, and radiology departments can adopt this technology to reduce the labor-intensive task of manual annotation and improve diagnostic efficiency. This will also allow healthcare providers to manage increasing patient loads without compromising the quality of care.

Global Market Size & Growth

• The global AI in healthcare market is projected to grow from $15.1 billion in 2023 to $187.95 billion by 2030, driven by the increasing demand for AI-driven technologies that improve diagnostic accuracy and reduce healthcare costs. The market is expected to grow at a compound annual growth rate (CAGR) of 37% during the forecast period.
• The medical imaging AI market, which includes AI models for diagnostic imaging like the one described here, is expected to expand from $1.06 billion in 2022 to $10.16 billion by 2030, with a CAGR of 33.2%. This growth is fueled by the rising prevalence of chronic diseases, the demand for faster and more accurate imaging analysis, and the shift towards value-based healthcare.
• Pulmonary embolism detection is part of the broader cardiovascular disease diagnostics market, which is also experiencing significant growth due to an aging global population and increasing awareness of the importance of early detection. The adoption of AI technologies in this area is expected to contribute to the growth of the market for cardiovascular diagnostics.

STATUS

• US provisional filed Sept. 2024
• TRL 4: Tested on over 2,000 cases from the RSNA Pulmonary Embolism CT Dataset (RSPECT), showing promising results in diagnostic accuracy and generalizability. 
• Seeking strategic partners for licensing.