Digital pathology: how medical imaging and AI are revolutionising the field


Digital pathology has become increasingly important over the last 10 years. In its simplest form, digital pathology involves taking a very high-resolution picture of a microscope slide containing a sample of tissue. This high-resolution picture allows pathologists to examine the sample in detail and look for a wide range of phenomena. For example, signs of infection, cancer or abnormal numbers or types of cells that could indicate disease. Finding disease signs early usually gives a higher chance of more successful treatments. 

Technological developments in the field of digital pathology have made it possible for clinicians to make primary diagnoses. Digital pathology makes it easier to store, view and organise medical images, making them easier to access and share. It also facilitates rapid engagement and collaboration (for example in multi-centre studies) between professionals. Second opinions can be easily and quickly solicited from clinicians in other locations, saving time and costs, and reducing the risks associated with physically transporting slides. Furthermore, organising images in a transparent and consistent way makes it easier for clinicians and scientists to satisfy regulatory requirements.

Crucially, the increasing prominence of digital pathology has provided access to specialist software applications such as automated image annotator analysis tools. These are used to assist in the interpretation and quantification of medical images. These tools employ sophisticated algorithms – often based on data from thousands of subjects. Computational algorithms, when correctly applied, are objective, highly accurate and can provide pathologists with more information to help make a diagnosis.

Computational analysis of digital pathology images

The use of digital pathology and, more specifically, automated image analysis, have increased substantially in recent years. However, several barriers to more widespread adoption still exist. These include:

  • limitations in technology (particularly within NHS departments)
  • cost of equipment
  • data storage and data protection concerns relating to patient confidentiality
  • a reluctance among pathologists to rely too heavily on modern tech and software.

Added to this, pathology departments also have to consider interoperability with existing systems, lifecycle management and ongoing costs.

However, the transition to digital pathology, as part of a wider move towards a digital society, is already underway. This is expedited by advances in whole slide imaging (WSI) technology, software applications, laboratory information management systems (LIMS) and improved high-speed networking capabilities. All of these have made it possible for digital techniques to be more easily integrated into pathology workflows.

A key part of such workflows is the integration of computer-based predictions and human annotations. However, there needs to be clear labelling to indicate whether such annotations are computer or human-generated. Whilst digital imaging and computational analysis cannot replace the role of a skilled pathologist, it can be used to support their work. Think of using computer-aided diagnostic techniques to provide a second opinion. This is often both faster and more accurate than an assessment carried out by the human eye.

Human input is vital to train AI systems

It’s important to remember that AI systems are only as good as the images and data they’ve been trained on. Care must therefore be taken, as they can easily get things wrong in situations that wouldn’t phase a human expert. For example, tissues are often stained with dyes to show certain features more clearly, and the stain intensity can vary wildly in different hospitals. This can confuse AI systems and could cause a misdiagnosis. Being able to visualise data and images across institutions will be vital in understanding this bias and identifying where more AI training is needed.

Another future challenge will be to quantify mathematically what has been ‘observed’ across the tissue or sample by the computer. This will be done using techniques such as spatial statistics, enabling the avoidance of bias and vague descriptive terms. For example, near a cancer tumour you might find a great number of blood vessels, or certain types of cells around the cancerous areas. A human might write “More blood vessels around the tumour than in the healthy cells”. Using spatial statistics, you can give a score to this type of phenomena, to accurately quantify them in an unbiased way. This will be essential when assessing large image-based collections to look for trends across populations.  It is also useful for monitoring disease over time, where many samples may have been gathered and need comparing.

In summary, the integration of digital pathology, annotation, AI and data visualisation software enables pathologists and clinicians to:
  • Accurately extract insights from large data sets
  • Collaborate with other professionals in a transparent way
  • Visualise and analyse data samples at speed
  • Improve quality control by helping identify outliers and bias early on
  • Utilise AI algorithms in the analysis of data sets and images
  • Improve accuracy and support clinicians in their diagnoses
  • Quantitate more accurately phenomena rather than using descriptive terms, allowing better assessment of the images
  • Improve tracking of disease conditions
  • Improve outcomes for patients

Keeping humans in the loop at all stages will be vital to ensure accuracy and safety for patients. Combining computational techniques with the power of human expert knowledge will provide better and faster diagnosis. This will save time, money and valuable resource, whilst enabling healthcare professionals to focus on other aspects of patient wellbeing.