Computer aided diagnosis for Rheumatic Heart Disease by AI applied to features extraction from echocardiography

Rheumatic Heart Disease (RHD) is a complication of rheumatic fever with higher incidence in developing countries. It causes inflammation and fibrosis of heart valves, among others, and can lead to severe heart damage, heart failure and death, it not treated in its early stages.

The Challenge

Process a large set of medical images, along with additional metadata and clinical information, efficiently and securely, to extract features that could be used to assist and even automate diagnosis.

The Solution

ATMOSPHERE developed an automatic method to assist its detection, with the aim of achieving an early diagnosis and improve patients' quality of life.

ATMOSPHERE developed an intercontinental infrastructure across Europe and Brazil in the frame of medical imaging processing, that offers a secure framework for deploying radiological imaging applications able to keep critical data encrypted and making them accessible only in trusted execution environments.

Materials and methods

Retrospective data from PROVAR (Programa de Rastreamento da VAlvopatia Reumatica) initiative in Brazil were used. This database was composed by 5600 echocardiography studies (5330 labeled as normal; 238 as borderline RHD; 32 as definite RHD). To compensate for that imbalance, the same number of pathological cases (borderline and definite) and normal cases were selected. Clinical endpoints were extracted from Doppler sequences in a fully-automatic manner by: frame splitting; differentiation in Doppler and anatomical frames by color inspection; color-based segmentation through k-means clustering; preprocessing and view classification using a Convolutional Neural Network; first- and second-order texture analysis and blood-flow velocity calculation; z-score features normalization; and features classification. Deep learning models were computed considering two classes for classification: healthy and pathological.

Furthermore, all sensitive data is processed using SGX containers, which protects them from memory reading-based attacks. It also includes an anonymization module responsible for removing potentially identifying information from medical images. Once critically sensitive information has been securely removed, images can leave the infrastructure deployed in Brazil and be stored on the European volumes. The GPU resources available in Europe are leveraged to perform the data processing. This infrastructure was used to set up a secure virtual environment to create an automatic classifier of echo-cardio images for the screening of Rheumatic Heart Disease (RHD). The application employed high-end computing resources and
ensured that sensitive data were only accessible within the region of origin.

Data storage and computing were performed on the ATMOSPHERE platform, capable of providing secure data storage even with unreliable cloud resources and implementing a self-managed computing back-end. Results. After a benchmark of classifiers, the best performance was obtained through logistic regression (75.6% of accuracy). The model provided a sensitivity of 78.79% and specificity of 72.73%.

Conclusions

This methodology has shown potential to help physicians in the diagnosis of RHD by performing an initial screening of cases, thus reducing the workload and improving the time-efficiency of health personnel while enhancing patient outcomes.

• Watch ATMOSPHERE webinar for more detailed info
• Download the Powerpoint with more info