Strategy & Solution

CTG developed a deep learning model to automatically segment lungs from CT imagery. To develop our automated lung segmentation model, we utilized a deep learning architecture known as U-Net. U-Nets are neural networks that have two main parts: a downsampling layer, which reduces the dimensionality of the input image, and an upsampling layer, which generates the segmentation mask. We customized this architecture using Tensorflow, a popular open-source framework for building and training machine learning models. One of the most important considerations in developing any machine learning model is selecting an appropriate loss function. In our case, we used a loss function called dice loss. This function measures the overlap between the predicted mask and the true mask, which allows us to determine how well our model is performing. However, one challenge in developing an automated lung segmentation model is that CT scans can come from different sources and may have different voxel spacing (the physical dimensions of a pixel) and slice thickness. To address this, we used a technique called resampling, which ensures that all scans have uniform voxel spacing and thickness. This helps improve the performance of our model and makes it more robust to variations in the data. To evaluate the performance of our model, we used a metric called the surface-dice score. This score measures the overlap between the predicted mask and the true mask along the edge or "surface" of the mask. We allowed a tolerance of 2mm, which means that differences less than that threshold are considered correct.