Today’s ML fact talks about how machine learning, in specific, deep learning helps in detecting tumors. Typically convolutional neural networks (CNNs) are used to detect tumors in MRI images. Two common ways that people use CNN structures are either to build classification models or to use architectures such as U-net and Segnet to do semantic segmentation.
For example, the work of Qi Zang et al  shows how to build a simple CNN based classifier that identifies malignant tumors. Another example is the work of Wang et al  where they train a classifier which outputs a heat map of a slide image. The heat map represents which part of the image belongs to the tumor and which does not. This type of classification can be effectively achieved using semantic segmentation as well.
Semantic segmentation involves pixel-wise segmentation which allows for detecting tumors of various shapes and sizes. This is possible because semantic segmentation gives us a pixel-wise map where the tumor is labeled differently compared to the background. An example of this is the work by Dong et al  where they use a U-net architecture to segment out brain tumors. Architectures like U net  have been used for other segmentation tasks as well such as cell segmentation.
All of this above analysis depends on quality data. Hence a standard labeled dataset for detecting tumors is needed to compare results. A benchmark dataset for this purpose is the BraTS dataset . The BraTs dataset contains MRI images which are labeled by neuroradiologists to reflect regions of tumors. More data is being continually added each year to this dataset.
There is immense scope for machine learning to help in solving some of the greater challenges in medicine and assisted therapy. As we come up with more ML facts we will highlight some of these new and interesting developments.