BS-Net is a new architecture designed to predict the severity of COVID-19 based on clinical data from different sources. This architecture uses four different blocks, which work together to estimate a six-valued score of the disease. This score is based on the interpretation of CXRs, which can be difficult and produce inter-rater variability among radiologists.
How BS-Net Works
The input image is processed using a convolutional backbone known as ResNet-18. Then, segmentation is performed using a nested version of U-Net (U-Net++) to identify the region of the lungs. The alignment is estimated through the segmentation probability map, which is achieved through a spatial transformer network that can rotate, center, and zoom the lungs correctly. The multiresolution feature aligner produces input feature maps that are well focused on the specific area of interest. Eventually, the output of the FPN layer flows in a series of convolutional blocks to retrieve the output map, which is classified using a final Global Average Pooling layer and a SoftMax activation.
BS-Net uses a weakly-supervised approach for training due to the visual interpretation of disease signs on CXRs, where subtle findings or heavy lung impairment can produce unavoidable inter-rater variability among radiologists in assigning scores. The alignment block is pre-trained on a synthetic alignment dataset in a weakly-supervised setting using a Dice loss to help improve the classification accuracy of the model.
The Benefits of BS-Net
The architecture of BS-Net is designed specifically for COVID-19 severity prediction, making the scoring more accurate and reliable. Since it is based on clinical data from different sources, it can help reduce the workload of radiologists or doctors who need to interpret CXRs for COVID-19 diagnosis. The weakly-supervised approach used by BS-Net also helps to overcome the inter-rater variability caused by the visual interpretation of CXRs.
The Future of BS-Net
BS-Net is still relatively new and requires further testing and refinement before it can be widely adopted. However, it has already shown great potential for COVID-19 severity prediction, and there is potential to adapt it for other lung diseases. The architecture could also benefit from integration with other diagnostic tools, such as blood tests or respiratory function tests, to improve its accuracy and reliability. Overall, BS-Net is a promising development in the fight against COVID-19 and other lung diseases and will likely play a role in the future of disease diagnosis and treatment.
