Explainability in medical deep learning

Deep learning achieves superhuman performance in many medical tasks, but is rarely used in practice because of its lack of interpretability. Various sources of
error, such as biased data or adversarial attacks can, with little transparency, undermine the performance of such models. In medical applications these erroThe aim of this project is to make deep learning for medical imaging more interpretable. Current approaches typically operate in a black-box fashion, where an image is implicitly represented in a non-interpretable feature space and then mapped to a diagnosis. This makes it hard to debug, and ultimately, to trust a model. An alternative approach would be to learn to describe an image and use these descriptions as rationales for the diagnosis. Inradiology, medical images
are accompanied by radiology reports, in which physicians describe what they see on the scan and make a diagnosis based on their findings. As such,
these reports can be used as training labels for both the diagnosis and the explanations that should be yielded by our model. The end goal is a framework that takes a medical scan as input and then provides some target information, which it explains by describing, and potentially locating, different findings in the image. For example, the input could be a chest X-ray and the output could be "acute pneumonia" (target information) because "consolidation within the medial aspect of the right middle lobe obliterating the right cardiac margin"(explanation). rs can be life threatening and the ensuing ethical, legal, and regulatory challenges therefore often prohibit the use of these approaches. If medical deep learning methods could overcome the obstacle of interpretability, they could be used more widely and improve patient wellbeing.This approach would lead to higher interpretability and, in addition, has the potential to improve the diagnostic performance of the model. Further, decomposing images into elementary building blocks (i.e. radiographic findings) can make it easier to incorporate background knowledge and is well suited for small data problems.Aims and objectives: Extracting findings from clinical reports: Clinical reports (e.g. radiology reports) provide rich descriptions of medical images. Besides a diagnosis, they typically also contain other findings, which explain the diagnosis. The goal would be to extract information from the reports in a structured format, in order to obtain diagnosis and explanation labels. An auxiliary task could be to predict the conclusion section of radiology reports (containing diagnosis and follow-up recommendations) from the findings section (containing descriptive radiographic findings), without the use of images. Large publicly available datasets of medical images and their corresponding reports, such as the MIMIC-CXR Database (consists of 227,835 radiographic studies) or PADCHEST, can be used for this task.Scene understanding of medical images: The labels generated in the previous aim can be used to train a model to take a medical image as input, and output both a diagnosis and explanations that justify the diagnosis. Initially, the aim is to yield

In the same way that bioinformatics has transformed genomic research and clinical practice, health data science will have a dramatic and lasting impact upon the broader fields of medical research, population health, and healthcare delivery. The beneficiaries of the proposed training programme, and of the research that it delivers and enables, will include academia, industry, healthcare, and the broader UK economy.

Academia: Graduates of the training programme will be well placed to start their post-doctoral careers in leading academic institutions, engaging in high-impact multi-disciplinary research, helping to build training and research capacity, sharing their experience within the wider academic community.

Industry: Partner organisations will benefit from close collaboration with leading researchers, from the joint exploration of research priorities, and from the commercialisation of arising intellectual property. Other organisations will benefit from the availability of highly-qualified graduates with skills in big health data analytics.

Healthcare: Healthcare organisations and patients will benefit from the results of enabled and accelerated health research, leading to new treatments and technologies, and an improved ability to identify and evaluate potential improvements in practice through the analysis of real-world health data.

Economy: The life sciences sector is a key component of the UK economy. The programme will provide partner companies with direct access to leading-edge research. Graduates of the programme will be well-qualified to contribute to economic growth - supporting health research and the development of new products and services - and will be able to inform policy and decision making at organisational, regional, and national levels.