In recent years, deep learning techniques have become the state of the art in computer vision, making artificial intelligence a growing reality in many areas of modern life. The impact can already be seen in manufacturing-led fields such as driverless cars and robotic assistants.
The challenge for deep learning techniques is having access to sufficient data from which to develop the algorithms. Deep learning focuses on the development of computer programmes that automatically extend and adapt their knowledge when exposed to new data: the more data that is available, the more accurate the decision-making.
Deep learning techniques have been used successfully in general image recognition tasks, with the development of a specific neural network sub-type (convolutional neural networks) becoming the de facto standard. Unfortunately, this has so far failed to transfer to the more specific field of medical imaging. In the main, this is because insufficient quantities of images are available to develop the complex set of algorithms needed for accurate decision-making.
Jeroen van der Laak
Solving complex diagnostic challenges
However, there is potential for a breakthrough from another branch of diagnostics – histopathology, the study of diseased tissue. This is a good candidate for deep learning techniques because it produces ‘big data’ (one scanned tissue section comprising many gigabytes of data). A great example of building a large repository of image data for deep learning is the recently announced collaboration between Philips and LabPON. They plan to create a digital database of massive aggregated sets of annotated pathology images and big data.
There are actually many examples in histopathology that could successfully be solved by deep learning. This article deals with two – finding breast cancer metastases in lymph nodes; and its use in the detection of prostate cancer in biopsy specimens.
Standard prostate cancer biopsy procedure gathers 8–12 separate, random, biopsies, guided by ultrasound, and resulting in multiple slides. However, the majority typically do not contain cancer. Studies at Radboud already show that deep learning techniques have the potential to significantly streamline the histopathological analysis if the algorithm can automatically exclude the normal slides without rejecting any containing cancer. It is estimated that up to 32% of slides could be confidently excluded in this way from further review.1
Further, due to the subjective nature of microscopic examination, variability of results is accepted, not just between hospitals but between colleagues. A consistent and robust algorithm removes that variability. Another challenge is the risk of a false positive. This can be a particular problem with tissue at the edges of the biopsy sample, which can appear abnormal just because it has been damaged. Over-diagnosis and over treatment of indolent cancers (slow growing) drains medical resources and can potentially subject patients to unnecessary and invasive procedures.
A further challenge therefore is not just to be able to determine whether cancer is present, but whether the cancer is indolent or aggressive. In diagnostic terms, the Gleason score refers to the way pathologists grade a prostate sample. They give it a number based on how much the cells in the cancerous tissue look like normal prostate tissue under the microscope. A subsequent stage for the studies at the university will see deep learning techniques used to determine whether a cancer will be slow growing or aggressive. This could lead to patients being safely offered more conservative follow up, such as active surveillance, rather than hospital-based treatment.
Computational pathology is a relatively new discipline within Radboud University, which was incorporated into its disruptive technology team – the Diagnostic Image Analysis Group (DIAG) –- less than four years ago. The objectives are clear – to have a significant impact on healthcare by bringing this technology into the clinic. The group is one of the first in the world to focus on the transfer of deep learning techniques to the histopathology laboratory and is forging diagnostic partnerships with industry, including Philips amongst others, to help make this a reality.
Radboud actively encourages collaboration with other deep learning groups worldwide and is willing to share information to maximise the data gathering for the algorithms. With this in mind, the DIAG computational pathology group initiated the two-year Camelyon Challenge,2 with last year 23 research groups contributing. They were asked to evaluate the robustness of new and existing algorithms for the automated detection of metastases in stained whole-slide images of lymph node sections.
For 2017, the challenge moves from just slide analysis to incorporating an assessment of patient outcomes, making it more relevant to the clinical setting. Results will be presented at this year’s International Symposium of Biomedical Imaging in Melbourne.
The DIAG computational pathology group are now in the final stages of developing a series of deep learning algorithms which could potentially be used in hospitals throughout Europe within the next five to ten years. These are specifically designed for the histopathologist to incorporate into diagnostic protocols, to improve the objectivity and efficiency of the diagnosis.
Impact on patient outcomes
The first addresses a specific diagnostic application: to detect cancer in lymph node samples, taken when determining the spread of breast cancer, for example. The presence of metastases in lymph nodes has therapeutic implications for breast cancer patients. However, while this is a clinically important investigation, it is tedious and time consuming when done manually by pathologists. An automated solution holds great promise for histopathologists as it could reduce their workload while at the same time removing diagnostic subjectivity and even increase accuracy.
This algorithm was first demonstrated and evaluated as part of the 2016 Camelyon Challenge. The DIAG computational pathology group expect to have reached sufficient accuracy for this diagnostic application to be ‘ready for use’ within the next 12 months. It would then, of course, need to be submitted for CE Mark and FDA approvals before being made available to hospitals, possibly within a further one to two years.
Radboud is also actively data gathering and computing for the far more complex requirement of prostate cancer detection from biopsy samples. While development time is longer, the university is confident this could be ready for hospital trials within two years (and then submitted for regulatory approvals).
The impact on patients could be life-changing, with early and more accurate diagnosis likely to lead to improved outcomes and more targeted treatment plans. Hospital budget holders will welcome the delivery of more efficient use of resources. However, the other main beneficiary is likely to be pathologists themselves. While in other areas, automating processes within the core laboratory often causes initial concern over job security and staffing levels, this is less likely to be the case with histopathologists. They correctly recognise that deep learning is far more likely to enhance their role, not undermine it.
Demand for new diagnostic roadmap
Due to the increase in cancer detection and the move towards more patient-specific treatment options, the diagnosis and grading of cancer has become increasingly complex. Pathologists have to review an increasing number of slides, as well as additional immune-histochemical stains, before reaching a subjective diagnosis and indicating appropriate treatment. More-over, there is an increase in the number of quantitative parameters they have to extract for commonly used grading systems.
In digital histopathology one whole slide image typically contains trillions of pixels from which hundreds of examples of cancerous glands (in the case of prostate or breast cancer) can be extracted by a computational pathology system. Further, the expansion of prostate-specific antigen testing in the past 20 years has exponentially increased the number of prostate biopsy sections. This continues to place an additional burden both on lab workloads and healthcare budgets.
The histopathologist may be regarded as the cancer diagnostic expert but today’s profession faces excessive workload pressures. Diagnostic delays and potential errors are therefore a daily concern. Pathologists have increasingly been looking for a new diagnostic road map; and their professional integrity has led them to actively collaborate with computer science experts like the group at Radboud University.
The growth in digital pathology imaging for routine diagnosis has signalled the possibility of applying these image analysis techniques to the examination and quantification of more complex slides. With a plethora of whole sample slides available globally for scanning, deep learning techniques offer the way ahead.
Resolving uncertainty
First and foremost, the practical application of deep learning for pathology would resolve uncertainty in the diagnosis of cancers. Diagnostic partnerships with industry would therefore enable Radboud to create practical and patient-focused applications that become part of the routine workflow and diagnosis protocols of the histopathologist.
This would be especially important in the challenging areas of prostate and breast cancer diagnosis, where over treatment and over diagnosis increasingly place a burden on both the patient and medical resources.
Of course, the potential for applying deep learning techniques is wider than clinical diagnostics. For example, they could be used to quickly analyse the data gathered in even the largest clinical trials to extract relevant case studies. Alternatively, they could automatically identify disease states or patterns to help determine future healthcare needs. Further, the technique could readily be applied to immunohistochemistry, which might be of interest when researching the efficacy of drugs or the expression of genes.
Perhaps the words of the conceptual artist Joachim Schmid, notable for his disruptive techniques, best expresses the creative potential of deep learning when he states: ‘there are no meaningful images. Meanings are created outside the image’.
References
1 Litjens G et al. Deep learning as a tool for increased accuracy and efficiency of histopathological diagnosis. Sci Rep 2016;6:26286.
2 CAMELYON 17. https://camelyon17.grand-challenge.org (accessed March 2017).
