What We Do
Histopathology is a central component of the diagnostic pipeline. Insufficient stuffing of histopathologists is a global problem and causes delays in the delivery of diagnosis which is often of decreased quality. In the UK, for instance, more than 13,000 cancer patients are waiting for more than 2 months to start treatment at any given time. In oncology these shortcomings severely affect the patients’ final outcome and quality of life.
In DeepMed IO we develop a modular AI-powered decision support system for enabling histopathologists to perform certain diagnostic tasks faster and with greater accuracy, initially with two modules, (a) pan-cancer metastasis detection on lymph nodes and (b) standardisation of breast cancer grading based on the Nottingham Grading System
Metastasis Detection on Lymph Nodes
Detection of metastases in Hematoxylin & Eosin stained slides of lymph node sections is of high clinical relevance but requires large amounts of reading time from histopathologists. We are currently developing a deep learning powered system for pan-cancer automatic detection of metastatic regions on lymph-node sections. This system will serve as an augmentation tool for pathologists increasing diagnostic speed and accuracy.
In December 2017 we received funding from the NHS Small Business Research Initiative to develop a proof-of-concept system for detecting metastatic regions on lymph node Hematoxylin & Eosin stained Whole Slide Images from breast cancer patients.
Our α-prototype reached an average of 97% accuracy on blind samples and achieved 53% decrease of diagnostic time (2.1 times faster, p-value=0.01318) and a concurrent 4.6% increase in diagnostic accuracy coming from detection of previously missed micrometastases, in a pilot clinical evaluation carried out by six experienced histopathologists(June 2018).
Having completed the feasibility stage of our project we quickly move forward to produce predictive models for melanoma, as well as lung, colon and prostate cancers which, along with breast account for more than 50% of all cancers globally.
Upon completion, our metastasis detection system will be embedded in a modular deep-learning powered computer aided diagnosis tool for digital pathology ready to be deployed in hospitals worldwide. The metastasis detection module along with other deep learning powered modules will automate a big range of the histopathologists’ time-consuming and laborious daily diagnostic routine and will also provide novel tools for predicting therapeutic responses and recurrence probability.
Deep-learning powered decision support for breast cancer grading
Breast cancer is a heterogeneous disease with varied morphological appearances, and response to therapy. Current routine clinical management relies on the availability of robust clinical and pathological prognostic and predictive factors, one of the best being histological grade, which represents the morphological assessment of tumor biological characteristics and has been shown to be able to generate important information related to the clinical behavior of breast cancers. The Nottingham Grading System (NGS), is the grading system recommended by various professional bodies internationally namely the World Health Organization, the American Joint Committee on Cancer, the European Union, and the Royal College of Pathologists.
The NGS places an emphasis on the percentage of tubule formation on the tumor, the mitotic count and the degree of nuclear pleomorphism, arriving at a ‘combined histological grade. These characteristics are determined by visual examination of tumor sections by expert histopathologists over a microscope. This practice presents two problems. The first one is that the histological grade is subjectively determined, where the main sources of variance are the histopathologist’s experience and fatigue levels as well as the fact that the tissue section is only partly examined. The subjectivity problem has been demonstrated by numerous cross-center studies where the overall average cross-expert correlation was found to be approximately 50%. To solve this problem we use deep-learning in a two-way approach. In the first we train our models to efficiently detect cell-nuclei, and then perform accurate mitotic counting and pleomorphism characterisation. In the second, the models are trained to accurately quantify the presence of tubules throughput the whole tumor tissue area.