When working with AI, doctors find it useful to see a region-of-interest, or an indication of which parts of the image the algorithm relies on most strongly while reaching a conclusion. This enables them to ‘see through the algorithm’s eyes’. Such area markers or heatmaps provide clinical users with visual cues that could make it clearer whether to accept or reject a chest x-ray finding detected by AI.

Algorithm interpretability is an active area of deep learning research — and a focus area for Qure.ai. The current visualization methods can be broadly classified into 2 categories: perturbation-based visualizations and backpropagation-based visualizations. We’ve experimented with these methods and put together a blog series on how they work with medical images, using chest X-rays as an example.

Read Part 1 and Part 2 of the blog post series, or the Auntminnie article about our RSNA presentation.

Head CT is the standard initial imaging study for patients with head trauma or stroke symptoms. We performed a study to validate the accuracy (versus 3 radiologists) of a set of deep learning algorithms that were trained to detect head CT abnormalities requiring urgent attention. The trained algorithms detect five kinds of intracranial hemorrhages (ICH) namely intraparenchymal (IPH), intraventricular (IVH), subdural (SDH), extradural (EDH) and subarachnoid (SAH), and skull fractures. They also detect mass effect and midline shift, both used as indicators of severity of the brain injury.

As part of the study, we’ve made a large head CT scan dataset, including 3 radiologist reads, available for public download in partnership with CARING, so that others can use it to develop and benchmark new methods.

Read about the study or explore the dataset.

Building accurate deep learning models requires large amounts of well-labelled training data. For radiology scans, using the accompanying clinical reports as ground truth is the most scalable way to supply classification algorithms with the data that they need to achieve high accuracy.

However, these reports are typically written as free-form text rather than in a structured format. Developing smart natural language processing (NLP) algorithms to tap into the expertise contained in radiology reports reports is important because it allows us to create ground truth labels at scale (hundreds of thousands to millions of scans).

Rule-based NLP systems use a list of manually created rules to parse the unorganized content and structure it. Machine Learning (ML) based NLP systems, on the other hand, automatically generate the rules when trained on a large annotated dataset. We developed a custom dictionary-based NLP method to read radiology reports, and compared it with machine-learning approaches.

Read more on the blog