For over a century, the study of tissue samples under a microscope has been a cornerstone in diagnosing cancer. Pathologists use a technique called Hematoxylin and Eosin (H&E) staining, which highlights cell structures by coloring the cell nuclei dark purple and other parts pink. This method makes it easier for doctors to examine tissues and identify cancerous cells.
In recent years, digital pathology has gained prominence with the development of advanced scanning devices. These scanners create digital versions of the stained slides that can be up to 5 gigabytes in size, equivalent to approximately 1250 high-resolution photographs. This large size indicates that the digital slide contains vast detail. Such detailed digital slides allow pathologists to zoom in and examine specific tissue areas without losing clarity, like looking through a microscope. These high-resolution images can contain a massive amount of data that clearly shows tiny cellular structures. The digitisation of these slides opens the door for artificial intelligence (AI) to assist pathologists in making faster and more accurate diagnoses.
The role of AI in pathology
AI, especially through deep learning algorithms, has demonstrated significant potential in analysing digital pathology images. By identifying patterns within these images, AI can assist in classifying various types of cancer and forecasting patient outcomes. For example, AI systems have achieved high accuracy in grading prostate cancer, comparable to the expertise of seasoned pathologists.
This technology offers the promise of quicker and more precise diagnoses, leading to more effective treatment plans and enhanced patient outcomes. AI supports doctors by highlighting crucial areas within tissue samples, enabling pathologists to concentrate on essential details. This collaboration between AI and pathologists improves both the speed and accuracy of cancer diagnoses.
Challenges in implementing AI
Despite its potential, the implementation of AI in clinical settings faces several challenges. Deep learning models require extensive training data to function effectively. However, acquiring sufficient data from a single medical centre can be difficult due to the limited number of cases available.
In addition, variations in tissue preparation and staining techniques across different hospitals lead to inconsistencies in slide appearances. These discrepancies create difficulties for AI models, which may struggle to generalise when exposed to slides from various sources. While pathologists can adapt to these differences, AI models often tend to memorise specific patterns rather than learning to generalise, which limits their effectiveness.
For instance, consider an autonomous vehicle trained to recognise road objects such as stop signs and pedestrians. If the vehicle can only detect these objects under ideal conditions, like daylight, its functionality is compromised in less optimal situations, such as at night or in fog. Similarly, AI models in pathology must be designed to perform reliably under diverse conditions to enhance their effectiveness and dependability.
Overcoming the challenges with advanced techniques
To address these challenges, researchers are developing techniques to improve AI models’ ability to work with data from multiple centers. One approach is domain adaptation, which involves training AI models to recognize and adjust to variations in data. Adversarial networks, an AI model, align features from different data sources, helping the AI focus on relevant details.
Another approach is colour normalisation, which adjusts the colors in images to create consistency across slides from different laboratories. While this method improves AI performance, it has its limitations. It often requires manual selection of reference images and may not fully eliminate the model’s ability to identify the original data source, such as recognising the specific medical centre from which the data originated.
Introducing adversarial fourier-based domain adaptation
Researchers at the University of British Columbia’s AI in Medicine Lab (AIM Lab) have developed a new technique called Adversarial Fourier-based Domain Adaptation (AIDA). This method leverages the Fourier transform to enhance AI models’ ability to identify key features in tissue samples, improving their sensitivity to colour and brightness variations.
AIDA incorporates an FFT (Fast Fourier Transform) Enhancer module, which decomposes images into simpler frequency-based patterns rather than relying solely on pixel values. By converting image data into the frequency domain, the FFT enables AI models to focus on critical features, such as shapes and structures within the tissue, which are essential for accurate cancer diagnosis. This approach enhances the model’s ability to generalise across different medical centres, thanks to its simplicity and efficiency.
Testing and results
Researchers tested the AIDA approach on a wide range of cancer samples, including ovarian, pleural, bladder, and breast cancers, from various hospitals. The results demonstrated that AIDA outperformed traditional methods, accurately identifying cancer subtypes and tumor regions. Expert pathologists validated the AI’s ability to recognize specific features related to each cancer type.
The superior performance of AIDA highlights its potential to improve the generalization of AI models when dealing with multi-center histopathology datasets. By addressing the challenges of domain shift and color variations, AIDA can help ensure accurate diagnoses, regardless of where the samples are processed and analyzed.
Credit. npj Precision Oncology
Addressing health disparities with AI
Many remote or underserved areas may lack access to highly skilled pathologists and medical experts. This can lead to delayed or inaccurate diagnoses, impacting patient outcomes. With the help of AI in remote clinics, this technology can offer reliable diagnoses where specialists aren’t available. It can also help reduce errors and give a second opinion improving accuracy. Even in well-resourced environments, this is of immense value; however, it becomes an especially helpful tool in complex cases where human expertise may vary.
Conclusion
AI is revolutionizing the field of pathology by offering tools that enhance diagnostic accuracy and efficiency. Despite challenges, advances like AIDA pave the way for AI models to become more adaptable and reliable in real-world clinical settings regardless of location or economic status. By improving the generalisation of AI models, we can bridge the gap in cancer diagnosis and ensure patients receive timely and accurate care.
The integration of AI into pathology holds the potential to transform cancer diagnosis, leading to better health outcomes and saving more lives. As research progresses, AI will play an increasingly important role in supporting pathologists and improving patient care globally.
Some actionable recommendations
- Promote inclusive policy frameworks
- Global data collection & partnerships: Establish partnerships with hospitals and research centres in poorer, underserved regions to collect data. This ensures that AI systems do not solely reflect the conditions prevalent in wealthier areas. Governments and industry bodies should collaborate with international health organizations to gather diverse datasets for training AI models in cancer diagnosis, ensuring they represent different ethnicities, age groups, and socio-economic backgrounds.
- Stakeholder engagement: Governments and industry bodies should engage with diverse stakeholders, including those in marginalized communities.
- Establish global standards and ethical guidelines
- Monitoring and compliance: Set up an international body to monitor compliance with these standards.
- International standards: Collaborate with global health organizations like WHO to create standardized initiatives and ethical guidelines for AI in cancer diagnostics.
- Invest in education and capacity building
- Educational initiatives: Fund educational programs in developing countries to build local expertise in AI, ensuring these communities can participate and benefit from these AI advancements.
- Local research initiatives: Support local research initiatives in these regions to adapt AI tools.
- Reducing global digital divides: Encourage technology transfer programs that allow developing nations to access AI technologies and resources.
- Ensure equitable access to AI-driven cancer diagnostic tools
- Subsidized AI tools: Provide financial assistance to healthcare facilities in underserved areas to adopt AI-driven cancer diagnostic tools.
- Public-private partnerships: Foster collaborations between governments, tech companies, and NGOs to distribute AI diagnostic tools in regions with limited healthcare infrastructure.
🔬🧫🧪🔍🤓👩🔬🦠🔭📚
Journal reference
Asadi-Aghbolaghi, M., Darbandsari, A., Zhang, A., Contreras-Sanz, A., Boschman, J., Ahmadvand, P., … & Bashashati, A. (2024). Learning generalizable AI models for multi-center histopathology image classification. NPJ Precision Oncology, 8(1), 151. https://doi.org/10.1038/s41698-024-00652-4
