AURA Health Technologies is revolutionizing breast cancer surgical planning and navigation by integrating machine learning with advanced ultrasound technology. This innovation enhances surgeon capabilities, improves patient outcomes, and reduces healthcare costs by eliminating traditional localization methods. AURA's commitment to product development and collaborations with medical professionals is elevating the standard of breast cancer care.

The core functionality of Aura Health Technologies is the detection of biopsy markers, the segmentation of tumor margins and the visualization of the results intra-operatively and in real time. The high variance and noise in ultrasound data creates the need to collect large amounts of data from many different patients in order to achieve the quality and reliability of model predictions necessary for use in the medical field. 

‍

1. Aura has now collected and annotated intra-operative ultrasound data from more than 250 breast cancer patients. At present, there are approximately 300,000 annotated ultrasound frames available, although these are very unevenly distributed among the individual patients. There are individual patients with > 10,000 annotated frames, and others with only < 100 annotated frames. This results in a strong bias in the training and validation data, which can have a negative effect on the generalization quality of the trained models. In addition, the large amount of data leads to very long training times, which means there is less time to optimize the hyperparameters, which in turn has a negative effect on the model quality. Reducing the data set by randomly selecting individual frames does not lead to success in this case, as, especially in the case of patients with a large number of annotated frames, a large number of very similar (almost identical) frames and only a few marginal cases were included. The consideration of these marginal cases is of great importance for good model quality, but would possibly be lost through random selection. The Lighlty software allows us to significantly reduce the size of our data set while maintaining high diversity in the data. 

‍

2. Labeling medical images is a critical yet costly aspect in healthcare technology. Recognizing this challenge, AURA Health Technologies was searching for a solution to minimize their labeling expenses and optimize their workflow. The traditional process of labeling is not only time-consuming but also requires significant financial resources and expert input. This situation posed a barrier to efficiently scaling their operations and advancing their innovative ultrasound technology for breast cancer treatment. Therefore, AURA Health was driven to find a more streamlined and cost-effective approach to handle medical image labeling, which is vital for the accuracy and effectiveness of its AI-driven tools.

1. Reduction of the Dataset

During our development, we investigated the effects of different dataset sizes on the training time per model and the resulting accuracy of the predictions on a representative validation dataset. To do this, we used the Lightly software to select subsets of different sizes while maximizing the diversity of the data in each case.  The results are summarized in the charts below. 

2. Reduction of future labeling effort through active learning
   

Using Lightly, Aura Health was able to reduce the dataset sizes by up to 80% without a reduction of model accuracy.