Frontiers in Histopathology – Where to Next?

Frontiers in Histopathology – Where to Next?

Author: Eva Hanlon 

Histopathology, the microscopic examination of tissues for the diagnosis of disease, has been a cornerstone of medicine for centuries.¹ Until recent years, abnormal tissue samples have been studied on glass slides with a light microscope by a pathologist, forming on its basis an expert, albeit subjective, assessment and diagnosis.² Times are fast changing however, with histopathology now hot on the heels of other medical fields transitioning towards automation, quantification, and objective evaluation.³ Without further aside then, what does this transformation look like for histopathology? 

Providing a whole new lens is digital pathology, where glass slides are scanned to produce high resolution digital images for viewing on a computer screen.¹ Increasingly armed with sophisticated artificial intelligence (AI) algorithms, these systems are further able to highlight and quantify areas of interest.⁴ Defined as the simulation of human intelligence by machines for autonomous task completion⁵, AI in reality now exceeds our own intelligence in several regards, able to process and retain much larger data sets, spot non-linear patterns, and integrate multi-modal data formats - all with superior speed, acuity, and sensitivity. Within AI lies the subset of machine learning (ML), where an algorithm ‘learns’ from extensive input data to subsequently make informed predictions.⁶ In the context of histopathology, these algorithms enable precise feature quantification, detection of subtle abnormalities likely missed with the human eye, and heightened accuracy of and consistency between judgements.¹ The digitisation of pathology also presents important logistical benefits, facilitating case sharing and collaboration between professionals worldwide, remote viewing of cases (telepathology), and significantly reduced risk of slide loss and/or damage.^1,7 Pathologists also appear to be welcoming the change, with 90% and 60% feeling comfortable making a digital primary diagnosis with or without access to glass slides respectively, in one 2020 study.⁸

One area where computational pathology is proving especially transformative is cancer, diagnosing malignancies with an accuracy rivalling and sometimes exceeding that of pathologists. For instance, one AI algorithm demonstrated 87% accuracy in distinguishing slides showing normal, benign, in situ, or invasive breast carcinoma lesions, outstripping the 80% awarded to pathologists on the same task.⁹ Elsewhere deep learning (DL) models, utilising artificial neural networks, are even showing their ability to predict metastasis. One such algorithm, trained on a mere 118 haematoxylin and eosin (H&E) diagnostic images from lung cancer patients, distinguished eventual development of brain metastases with 87% accuracy, again far exceeding the comparative 57% exhibited by specialists.¹⁰

Although remarkable, such performances may in some, instil fears of expert redundancy – after all, what is the need for a pathologist if anything they can do, AI can do better?

The answer lies in their synergy.

ML models can only make judgements based on their prior learning experiences, which in themselves rely on datasets often limited in size and quality.¹¹ Rare and more complex cases thus invariably still require assessment by a professional, with AI simply leaving them more time to do so. Indeed, in the aforementioned study considering breast cancer diagnosis, pathologists aided by the AI programme improved their accuracy to 88%, reiterating the clear potential for collaboration between these two ‘experts’ in their own right.⁹ 

Data quality concerns however, are from trivial, as simply put: what you get out of a ML model will only ever be as good as what you put in. As well as risking decision inaccuracy, poor or limited input data may also introduce implicit bias.¹¹ One potential strategy to overcome this challenge is data sharing between institutions, although this is accompanied by its own complications regarding data protection and heterogeneity.¹² Transfer learning is likely a more promising approach, where algorithms are fed existing knowledge from different but closely related models.² The complexity of implementing computational pathology in cancer care also doesn’t end there. Wielding such heightened sensitivity, ML models may become a double-edged sword, running the risk of ‘overdiagnosis’ whereby detection of in situ/low grade cancers may lead to the recommendation of overly-drastic treatments.  Caution in treading this fine line is therefore critical, as warned by one clinician “diagnoses of early stage cancer made using ML algorithms will undoubtedly be more consistent and replicable…but they won’t necessarily be closer to the truth – determining which tumours are destined to cause symptoms or death.”¹³

One potential road to middle ground is the multi-omics approach, combining histopathological image analysis with genomic data, in an attempt to merge genotype and phenotype to better approximate overall risk.¹⁴ Such an approach is not limited to solely genomics either, with the potential to integrate epigenomics, proteomics, transcriptomics, and metabolomics data at the single-cell level.  Even non-omics data, such as electronic health records can be utilised, allowing algorithms to ‘learn’ from patient outcomes to subsequently recognise groups unlikely to benefit from diagnosis.¹² With such an intricate picture of an individual’s disease, highly informed decisions could then be made regarding patient diagnosis, prognosis and treatment – fast beckoning in the gold standard of personalised medicine. Continuing with the example of cancer, as well as distinguishing cancer subtypes from H&E slides, DL models are now also capable of predicting cancer-relevant gene alterations and microsatellite instability from images.¹⁴ Image2TMB for example, predicts lung cancer tumour mutational burden from H&E slides with accuracy equivalent to large panel sequencing, providing an invaluable predictor of response to immunotherapy.⁵ Given the severe toxicities associated with these treatments, tools such as Image2TMB could be instrumental to obtain an informed risk versus benefit assessment. Whilst the integration of such heterogeneous data and the allocation of appropriate weight to each variable is inevitably complex, combining digital pathology with rapidly expanding omics promises a gold mine of information, one we could only be unwise not to seek.   

Every tumour however, exists in four dimensions: three spatial dimensions (x, y, z), and the fourth dimension of time; any 2D representation therefore inevitably falls short of representing the disease process in its entirety.¹⁵ Accordingly, histopathology is now seeking to transcend the use of 2D sections in favour of multi-dimensional analyses. 3D reconstructions of H&E sections for example, incorporate the z-axis to enable the pathologist to effectively ‘move through’ a tissue.¹⁵ Further leaving behind the ‘snapshot’ captured by classical histopathology, DL models can now analyse the genome/transcriptome of cells whilst maintaining their spatial relationships, deciphering tumour, stromal, and immune cell localisation within the tumour microenvironment. ‘DNAseq’ is one such spatial genomics technology, which in addition to conventional morphological features, sheds light on tumour progression by identifying distinct tumour clones.¹⁶ Spatial transcriptomics unravels other complexities, such as intra-tumoural heterogeneity of biomarker expression.¹⁷ This pioneering integration of state-of-the-art microscopy, digital pathology, and spatial omics in the third dimension, promises to revolutionise our understanding of and approach to cancer. The road from preclinical exploration to clinical implementation however, remains rocky, as procedures continue to prove expensive and protocols incompatible with current workflows.¹⁵  

It is also worth noting that with the increasing complexity of DL models, their decision-making process becomes increasingly opaque or ‘black box’, obstructing interpretation by pathologists.^2,11 AI will therefore only realise its full potential in the field of histopathology when it presents itself as ‘explainable AI’, where efforts such as feature visualisation and saliency maps make its processes transparent for the user.¹⁸ Furthermore even with such precision and the extraction of human error, there nonetheless remains room for misinterpretation with AI-based histopathology. This inevitably begs the question – who is accountable for such an error if the guilty party is a trained AI system? After all, machines are not required to take responsibility for their actions by law.  Debates are ongoing as to whether AI “fits within existing legal categories, or whether a new category with its special features and implications should be developed”.¹⁹ Either way, the need for timely decisions is critical, given that a simple misjudgement in medicine, made by man or machine, can bear potentially devastating consequences. 

Although here discussed in the context of cancer, these rapidly advancing technologies are poised to transform histopathology in all human disease, with the potential to massively enhance diagnostic accuracy, improve patient outcomes, and further enable the paradigm shift towards personalised medicine.¹ Their translation from bench to bedside however, is proving a challenge, impeded by infrastructural obstacles, a lack of standardisation and quality control, and discordance with established workflows.^2,15 If the ready embrace of molecular diagnostics is anything to go by however, the field will rise to the occasion. As the horizons of histopathology continue to expand, only one question persists – where to next?

Bibliography:

1 Menditti, D. Exploring the Future of Histopathology: Innovations and Breakthroughs. J Cytol Histol 14, 708 (2023). 

2 Bahadir, C. D. et al. Artificial intelligence applications in histopathology. Nature Reviews Electrical Engineering 1, 93-108 (2024). h
ttps://doi.org:10.1038/s44287-023-00012-7

3 Mubarak, M. Move from Traditional Histopathology to Digital and Computational Pathology: Are we Ready? Indian J Nephrol 32, 414-415 (2022). h
ttps://doi.org:10.4103/ijn.IJN_508_20

4 Niazi, M. K. K., Parwani, A. V. & Gurcan, M. N. Digital pathology and artificial intelligence. Lancet Oncol 20, e253-e261 (2019). h
ttps://doi.org:10.1016/s1470-2045(19)30154-8

5 Tran, K. A. et al. Deep learning in cancer diagnosis, prognosis and treatment selection. Genome Med 13, 152 (2021). h
ttps://doi.org:10.1186/s13073-021-00968-x

6 Kann, B. H., Hosny, A. & Aerts, H. Artificial intelligence for clinical oncology. Cancer Cell 39, 916-927 (2021). h
ttps://doi.org:10.1016/j.ccell.2021.04.002

7 Vielh, M. The Evolution of Cytology and Histology in Medical Research. J Cytol Histol 15, 727 (2024). 

8 Hanna, M. G. et al. Integrating digital pathology into clinical practice. Mod Pathol 35, 152-164 (2022). h
ttps://doi.org:10.1038/s41379-021-00929-0

9 Polónia, A. et al. Artificial Intelligence Improves the Accuracy in Histologic Classification of Breast Lesions. Am J Clin Pathol 155, 527-536 (2021). h
ttps://doi.org:10.1093/ajcp/aqaa151

10 Zhou, H. et al. AI-guided histopathology predicts brain metastasis in lung cancer patients. J Pathol 263, 89-98 (2024). h
ttps://doi.org:10.1002/path.6263

11 Asif, A. et al. Unleashing the potential of AI for pathology: challenges and recommendations. J Pathol 260, 564-577 (2023). h
ttps://doi.org:10.1002/path.6168

12 He, X., Liu, X., Zuo, F., Shi, H. & Jing, J. Artificial intelligence-based multi-omics analysis fuels cancer precision medicine. Semin Cancer Biol 88, 187-200 (2023). h
ttps://doi.org:10.1016/j.semcancer.2022.12.009

13 Willyard, C. AI can help screen for cancer—but there’s a catch, (2023).

14 Cifci, D., Foersch, S. & Kather, J. N. Artificial intelligence to identify genetic alterations in conventional histopathology. J Pathol 257, 430-444 (2022). h
ttps://doi.org:10.1002/path.5898

15 Kramer, C., Vreeswijk, M., Thijssen, B., Bosse, T. & Wesseling, J. Beyond the snapshot: optimizing prognostication and prediction by moving from fixed to functional multidimensional cancer pathology. J Pathol 257, 403-412 (2022). h
ttps://doi.org:10.1002/path.5915

16 Zhao, T. et al. Spatial genomics enables multi-modal study of clonal heterogeneity in tissues. Nature 601, 85-91 (2022). h
ttps://doi.org:10.1038/s41586-021-04217-4

17 He, B. et al. Integrating spatial gene expression and breast tumour morphology via deep learning. Nature Biomedical Engineering 4, 827-834 (2020). h
ttps://doi.org:10.1038/s41551-020-0578-x

18 Klauschen, F., Dippel, J., Keyl, P. . Explainable Artificial Intelligence in Pathology. Pathology 45, 133-139 (2024). 

19 Naik, N. et al. Legal and Ethical Consideration in Artificial Intelligence in Healthcare: Who Takes Responsibility? Front Surg 9, 862322 (2022). https://doi.org:10.3389/fsurg.2022.862322