Histopathology Diagnosis Using Artificial Intelligence
Posted 26th June 2017 by Jane Williams
Pathological examination of a tissue section is considered to be the gold standard in medicine. It plays a critical and legal role in the diagnostic process.
Examination of tissue requires the biopsy/tissue sample to be surgically removed, placed into fixative to stabilise the tissue, processed overnight and then thinly sectioned onto a glass slide. In order to visualise the individual cell components under a microscope, the sections are stained with Haematoxylin and Eosin. Haematoxylin stains nuclei dark blue and eosin stains the cytoplasm and supporting structures pink.
Today, pathological diagnosis is performed by highly trained medical specialists. The training, over a period of at least 11 years, enables the histopathologist to recognise the specific pattern of the tissue under the microscope. The pathologist is responsible for the diagnostic and prognostic assessments, mainly based on the architecture and the individual cell morphology.
Over the last decade, computerised methods have rapidly evolved in digital pathology, with growing applications in research, consultations, quality assurance and as part of workflow improvements in laboratories. More recently, whole slide imaging is being explored through the ever-expanding field of machine learning and artificial intelligence. Automated recognition of pathology patterns in a whole slide image using machine learning is in its infancy, but it is showing promising potential to provide valuable assistance to the pathologist in daily practice.
Deep learning, a particularly effective variant of the machine learning paradigm, is highly effective at pixel-based distinction thus providing an effective platform for pattern recognition. It represents the state of the art approach for a variety of image analysis tasks includes detection and counting (e.g. mitotic events), segmentation (e.g. nuclei) and tissue classification (e.g. cancer vs non–cancerous). Irrespective of the task, Deep Learning learns to achieve it by analysing a set of labelled positive and negative examples.
By utilising machine learning on whole slide digital images, we have developed a proof of concept for automatic detection of breast cancer metastasis in a lymph node. Our early results show that metastatic breast tumour in a lymph node can be detected in microscopy images through artificial intelligence. Detection was shown to have a specificity of 89.6% and a sensitivity of 75%, which compares favourably to a human. Tumour samples less than 0.2mm or less than 200 cells were able to be detected.
Our method is based on a convolutional neural network (a variant of the Deep Learning approach) and specifically the Wide Res Net 50 residual network formulation. This model helps to overcome the vanishing gradient problems faced by comparable very deep networks.
In our approach, we first trained the CNNs, which were able to capture both fine level detail and high-level structure automatically, after training on a suitable set of examples. We then applied the trained deep model to partially overlapping patches from each whole slide image to create tumour prediction heatmaps.
Whilst deep learning develops powerful representations of the underlying data, the challenges such as selecting appropriate magnification at which to perform the analysis or classification, annotation accuracy in training datasets, and identifying a suitable training set containing information rich examples can not be underestimated. The cooperation between a deep learning expert and a pathologist is needed to translate the deep learning paradigm to digital pathology tasks.
Whilst it is unlikely that AI will entirely replace histopathologists, with time it will undoubtedly be in the position to provide increased efficiency and accuracy to the diagnosis, either by performing quality assurance or by providing a preliminary diagnosis.
Image provided courtesy of Ruth Salom.
Leave a Reply