A dinosaur’s tale. Morphology shouldn’t be relegated to the tail end of the cancer diagnostic process. Could AI-assisted feature mining could bring it back to the front end?

There is a fundamental relationship between morphology, behaviour and genetics. We cannot neglect any one of these to fully understand biological mechanisms in disease processes. The shape of a single well preserved tooth can tell the palaeontologist as much about a dinosaur’s behaviour than hundreds of sequenced genes. The microscopic patterns tumour cells make can also reveal not only the diagnosis, but also the genes altered and even potential targeted drug treatments.

I am a Consultant Histopathologist but started my career as a biologist before I studied medicine. My first fascination in biology centred around how form and function are intimately intertwined, and this continues to fascinate me in my daily clinical diagnostic work. Every form in nature is related to the function of that form; even if there is an obscure reason for it (like the peacock’s tail), there will always be a reason why evolution has crafted an organism the way it did.  And that rule holds true the deeper one looks at the structure of the organism, even at the microscopic level (particularly in cancer which is of course a rapidly adapting micro-evolutionary process, working at the molecular and cellular level).   During my former research as a cancer research biologist, I learnt to recognise the three main cell lines I grew in culture by their appearance under the microscope as if they were my own pets, despite them all being from the same cancer type with only a few genetic differences (I was studying the effects of different p53 mutations on apoptosis and radiation resistance in colorectal tumours). It is satisfying, but not surprising to me years later to be using those pattern recognition skills to distinguish cancer subtypes, and recognise genetic changes leading to changes in histopathological morphology. “Once a biologist always a biologist”, I still occasionally browse the (online) pages of journals like Science and Nature. These days the articles in the top journals (at least in those relating to my own interest of cancer biology) feature a bewildering array of new gene and molecule names and virtually nothing about microscopic morphology. This always strikes me as perverse, but scientific research has always followed trends and fashions relating to the funding supporting it and genes and molecules seem to be currently more “sexy” than the shapes and patterns of nature.
An apparently unrelated article about biological form and function however recently caught my eye, which relates to the tail of the “enigmatic” dinosaur Spinosaurus (https://www.nature.com/articles/s41586-020-2190-3) aegyptiacus [one might not immediately see the relevance of this to my argument but stick with me!]. A new fossil has shown that the tail shape of this dinosaur, previously thought to be a terrestrial predator, was unequivocally adapted to life in water. The tall dorsal spines on the beautifully preserved fossil tail clearly resembles that of a modern crocodiles or even a newts and creates a paddle shape for moving the organism through the water. 

Figure modified from Ibrahim et al 2020. The fossil spinosaurus tail has delicate dorsal struts nearly two feet long, giving it the profile of an oar. By the end of the tail, the bony bumps practically disappear, letting the tail’s tip undulate back and forth in a way that would propel the animal through water.

Morphological observations such as this require no need for novel genes to be identified (although genes are of course required to create these anatomical and physical characteristics), only knowledge of anatomy and behaviour. The same can be said about microscopic morphological changes associated with health and disease, and particularly with respect to cancer.  Cancer begins as an increased proliferation of cells which acquire a survival advantage through mutation of critical growth and metabolism-controlling genes, eventually achieving cellular immortality. Most simply explained, oncogenes are the growth accelerators, and tumour suppressor genes are the brakes on proliferation (sometimes these can also work through  inhibition of  programmed suicide mechanisms or signals relating to terminal differentiation). In an epithelial cancer (carcinoma), cells initially acquire the ability to lose contact with their neighbours, invade the basement membrane and infiltrate the mesenchyme, often undergoing prominent epithelial mesenchymal transition (EMT). The mesenchyme contains lymphatics and blood vessels (these vessels are the transport route to the metastatic site). Surviving in a foreign environment like vessels and lymph nodes relies on mutations in more genes, acquisition of growth factor independence as well as resistance to the host’s immune cells which are continually trying to attack and destroy them.  As the tumour mass increases in size it must recruit blood vessels (angiogenesis) and resist hypoxia to keep up with its disordered rapid growth. By this point most cancers will have become genomically unstable and have mutations in DNA repair mechanisms (for example mismatch repair deficiency leading to microsatellite instability).  Most if not all of these important “hallmarks of cancer” are visible to the histopathologist, either on standard H&E stains or with the addition of immunohistochemical stains to the same thin paraffin sections; whether or not the pathologist is able to reliably recognise and / or communicate the names and or significance of these morphological changes is a different question. Modern cancer therapeutic research is gradually discovering targeted treatments to each of these cancer hallmarks but current research efforts are almost exclusively based on identifying the underlying genes controlling these processes rather than the morphology (closely correlating with “tumour behaviour”) resulting from the underlying molecular changes.

The Hallmarks of Cancer modified from Hanahan and Weinberg 2011. Cell. Tumours with defects in one or more of these pathways also have microscopic morphological characteristics enabling their identification and potential therapeutic targeting.

As a trainee pathologist, I was amazed by the almost instinctual ability for an experienced histopathologist often to be able to comment on important features of a tumour before the microscope had even zoomed into a medium or high power objective.  Comments about “how fast this one is growing” or “how nasty this one looks” were often made at a magnification where individual cells were not even recognisable. This seems like magic to an untrained observer when they subsequently zoom in and confirm what they suspected when at a magnification where individual cells were not discernible, and is an apparent instinct honed by years of experience. There are of course epiphenomenona of mitotic or proliferation rate in tumours which are visible at low power plus a whole host of other unconsciously recognised features we are simply unable to put into words. We will often use vague descriptions our colleagues such as “this looks like colonic differentiation”, and would be embarrassed to use in a real report but are often confirmed after further stains, molecular markers and radiology. We are taught during our pathology training to recognise different types of tumours based on somewhat arbitrary textbook features, but even tumours of the same category and subtype from different patients can look as different as different human faces due to inherent cancer intratumoral heterogeneity.

It is not necessary to sequence the genes encoding proteins making up the skin, soft tissues and bones of these similar pairs of faces when we can use the pattern recognition skills of human or AI-trained neural networks. We can do the same with microscopic diagnoses of similar cancers driven by therapeutically targetable genes.

There is much excitement since the human genome project about identification of genes related to diseases and none more so than in cancer which will affect more than half of us at some point in our lives. Several genetic and molecular changes have now been identified which enable the use of targeted drugs, many of which can cure and/or considerably increase lifespan. Identification of a novel gene associated with rapid growth of a common cancer type could earn you a publication in a top journal like nature, but identification of a novel microscopic feature in tumour cells with the same prognostic significance will likely be relegated to the pages of an obscure histopathology journal.

Testing for targetable genetic mutations has become commonplace in our jobs as diagnostic pathologists. Our clinical colleagues increasingly prioritise these over making a diagnosis of a well defined morphological entity such as adenocarcinoma versus squamous carcinoma or even a more broad classification such as melanoma or sarcoma. This is because some treatments based on targeted mutations have now opened up curative treatment possibilities and prolonged lifespan for patients with metastatic disease that were previously only suitable for best supportive care. Histopathologists like myself obviously want the best possible outcomes for patients but are increasingly slaves to the modern oncological practice of “fishing for mutations” . Biopsy material is often scanty, and can be rapidly exhausted by “tumour agnostic genetic testing”. With increased understanding of morphology, features in the biopsy can be highly predictive or whether or not a particular gene will be altered. Random fishing trips can realistically become a targeted process, particularly with AI assistance.

As we learn to reliably recognise (with the help of AI) more microscopic patterns associated with treatable mutations we can stop wasting tissue on “oncology fishing trips”. The oncologist wants to identify mutations associated with drug response and sometimes requests multiple genetic tests on the same sample until the tissue is exhausted. The histopathologist can identify patterns associated with certain mutations, but many are subjective and not reliably communicated using current diagnostic terminology. With an AI “on his shoulder” the pathologist can help train neural networks to recognise recurring patterns and known prognostic histological features. With the addition of deep learning these datasets can be used to identify more subtle features many currently imperceptible to the pathologist to quantify and measure those most important in therapeutic prediction and prognosis.

As I have indicated already however, since genetic (and epigenetic) changes result in the changes in morphology we see in tumours under the microscope, should we be using a different “tissue-preserving” approach to help patients with cancer?

There has been a recent explosion in publications about artificial intelligence (AI) in histopathology since the increased introduction of slide scanning digital pathology technology (see my previous artlces on Digital Pathology, and AI/Phenomics). AI applied to digital pathology has multiple potential uses in laboratory workflow and beyond, which are summarised in the diagram below, including the increased automation of time consuming laborious tasks involving high levels of subjectivity and inter-observer variability such as tumour grading.

Recent AI algorithms are now available which already outperform expert pathologists in the recognition of certain types of cancer in biopsy specimens. Of even greater interest to me are recent data showing that AI neural networks can also be used to identify actionable mutations, even on H&E stained slides. These include prediction of Ki67 proliferation rate and hormone receptor status without need for antibody immunohistochemical (IHC) stains against these antigens. Algorithms have been trained on cancers showing microsatellite instability (MSI) and mismatch repair deficiency (dMMR) and subsequently shown to recognise features of MSI/dMMR on H&E stains of novel cases. The same approach has been used for an increasing number of tests and genetic alterations with high accuracies for some important prognostic and predictive markers. Why should we sequence the genetic sequence of a gene such as p53 when tumour cells show reliable and highly predictive morphological and immunohistochemical features of loss of function and the protein product can still be inactivated even when the gene sequence is intact? Some tests such as programmed cell death ligand (PDL1) which is used to predict immunotherapy response for lung cancer and an increasing number of other cancer types has always suffered from poor interobserver variability but early results using H&E deep learning algorithms also shows results as good as and in some cases better than trained pathologists. The advantage of using AI algorithms is that they will enable pathologists to reduce the time they take to perform laborious and subjective tasks like grading and feature quantification as well as increasing objectivity of their observations. In contrast to molecular (“fishing for mutations”) testing which can rapidly exhaust remaining material in a biopsy, AI-assisted pattern recognition does not consume tissue and unlimited analysis can be potentially performed on the same H&E WSI.

A lot of pathologists not surprisingly intimidated and sceptical about the idea of AI deep learning algorithms taking their jobs, and most I have spoken to hope that they will be retired before the technology gets a firm foothold in the UK. Pathologists felt the same way when IHC was introduced in the 1980’s but all have embraced this technology and interpretation of IHC results as one of their own skills. Unlike molecular testing which is generally often “outsourced” by pathologists and a text or numerical result received to incorporate or interpret into their final diagnoses, the results of AI algorithms will be immediately familiar to pathologists and our skills will be vital in determining their reliability and relevance to a particular case. Just like genetic testing, AI algorithms are critically dependent on the quality of the data entered into them, the output will be useless if relevant areas of the sample are properly annotated and clinically relevant. AI combined with pathologist (and genetic testing) will be much more powerful than either AI or pathologist alone. Pathologists will not be replaced by AI but pathologists who refuse to embrace new technology will be replaced by those who do. I hope to embrace and incorporate algorithms into my own work once we can introduce more widespread WSI. If we fail to do this it is very likely that clinicians will demand it is used (once improved outcomes related to more precise and objective diagnostics are realised) and the testing will have to be “outsourced” as we currently do for an increasing number of other specialised tests.

A parody of a Pathology AI system and pathologist working together. Will the future be that these tests are all analysed by a private company or will all public labs have their own digital pathology and AI systems using the same standards?

In summary I believe this technology will augment pathologist’ skills, breathing new life into our speciality and bringing us to the forefront of cancer diagnosis. The dinosaur’s tail is an reminder that the age old skills of careful examination and corroboration form and function are not yet dead and buried.

Leave a Reply

Your email address will not be published.