Research at the Harvard Biorobotics Lab 2018/2019:

Data processing, Model training, Statistical Evaluation

1) Cardiac Ultrasound Cone Segmentation:

The Harvard Biorobotics Lab at the 2019 ICRA Conference

This project was in collaboration with the Boston Children's Hospital during my time as a researcher at the Harvard Biorobotics Lab. For many deep learning tasks data preprocessing is an important step to achieve better classification and detection results. Especially in applying deep learning technology for medical applications the reduction of systemic bias is a key focus to ensure patient safety. A common source for medical data are cardiac ultrasound images. Often there is still additional information visible in the image, including static and dynamic elements like saturation bars, patient data or moving EKG lines. Those additional components can create a bias for deep learning algorithms to detect or classify the given image not based on the actual ultrasound data but the additional data. Therefore this project was aimed to create a image segmentation to create a binary mask in order to remove all non-ultrasound parts of the image. The segmented images were then used to create a classification to identify different structural cardiac diseases for children.

A sample input image with additional information, the automatically created binary mask and the resulting output image

2) Reducing the need for hand labeled data for deep learning models:

During the process of creating an ultrasound cone segmentation a new method to avoid large numbers of hand segmentation data was created. This allows to pretrain the existing model on less costly automatically created data and then improve the model on a small set of hand segmented data. The new method could be used for many deep learning applications since it drastically reduces the amount of labeled data that is necessary to train deep learning models.

This research project is still in the process of being published and the full results will be available soon