Editorial: Artificial Intelligence (AI), Digital Image Analysis, and the Future of Cancer Diagnosis and Prognosis

On October 8 2024, the Royal Swedish Academy of Sciences announced that the 2024 Nobel Prize in Physics was awarded to John J. Hopfield (Princeton University) and Geoffrey E. Hinton (University of Toronto) for their foundation research that has enabled machine learning with artificial neural networks, which have resulted in the current prevalence of and applications for artificial intelligence (AI) [1,2]. In the 1980s, Hopfield and Hinton worked on developing networks that differed from earlier computational methods because their networks could learn from examples and complex data [3,4]. In 1982, Hopfield, a theoretical biologist and physicist, published a critical paper on the computational properties of neural networks based on aspects of neurobiology and adapted to integrated circuits to produce a model with ‘content-addressable memory’ or ‘associative memory’ [3]. At the same time, a computer scientist, Geoffrey Hinton, created a tool to recognize and classify images to collectively describe systems with too many components to track individually [4]. In 1983, Hinton and colleagues published key research on the visual system and visual processing using a parallel algorithm [4]. In 2001, Hinton was awarded the Dirac Medal from the International Centre for Theoretical Physics (ICTP) [2]. In 2018, he received the AM Turing Award for contributions of lasting and major technical importance from the Association for Computing Machinery (ACM) [2]. It is of interest that this recent Nobel Prize decision has come at a time when applications of AI technology now involve all areas of human experience and endeavor, including social, artistic, scientific, and medical, and when the application of AI technology in image analysis, including in medical diagnostics, and infectious disease surveillance are increasing rapidly [5,6]. In November 2022, OpenAI launched the generative AI chatbot, ChatGPT, based on the GPT-3.5 and GPT-4 large language model (LLM), which is now widely and generally used [5]. However, the impact of AI systems, models, and technology in diagnostic imaging, mainly in histopathology, has steadily increased during the past decade [7].

Diagnostic histopathology is a specialty that relies on pattern recognition of the microscopy of tissue sections stained with histochemical stains to show cellular and non-cellular tissue components with patterns consistent with a tissue diagnosis [8]. During the 1970s and 1980s, immunohistochemistry and immunofluorescence methods using visual labels for polyclonal and monoclonal antibodies brought specificity to identifying the components of tissue sections when viewed by light microscopy [8]. It was quickly realized that the ‘by eye’ quantitation of immunostaining was too subjective, and early image analysis methods aimed to identify the staining intensity and the proportion of cells stained [8]. The development of targeted therapies in oncology, based on the expression levels of growth factors and their receptors, such as epidermal receptor (EGFR), resulted in the development of automated immunostaining methods (diagnostics) and improved quantitative image analysis [9,10]. However, image analysis technology was initially developed for one visual field at a time due to the computational challenges of analyzing a standard gigapixel slide consisting of tens of thousands of image tiles [7,11].

Digital pathology combines image capture with image analysis and uses digital tools to collect, analyze, and share diagnostic information [8,12]. For several decades, image capture and analysis have been performed directly via the microscope image [8,12]. Several medical device companies have recently developed whole-slide scanning devices, including companies with a history and expertise in manufacturing optical microscopes for diagnosis [12]. The digital image can be analyzed using high throughput algorithms, stored or shared [7]. In April 2017, the US Food and Drug Administration (FDA) permitted marketing authorization for the first whole slide imaging (WSI) system [13]. The Philips IntelliSite Pathology Solution (PIPS) (Philips Medical Systems Nederland BV) was developed to allow the review of digital surgical pathology biopsies routinely processed onto glass slides [13]. The PIPS was the first approval for a WSI system from the FDA Office of In Vitro Diagnostics and Radiological Health to streamline slide storage and retrieval and improve availability to pathologists and healthcare teams [13]. The PIPS relies on proprietary hardware and software to scan and digitize the histology images from histochemical and immunohistochemistry stained tissue sections on glass slides, with a resolution equivalent to a microscope objective of ×400 [13]. FDA approval was based on evaluated data from a clinical study of 1,992 surgical pathology cases that showed that clinical interpretations based on the PIPS images were comparable to those made by pathologist review of the glass slides [13,14].

Although image analysis and image analysis technology have become established in diagnostic histopathology, the application of AI modeling in diagnostic histopathology has just begun to gain momentum, driven by the increasing clinical burden of chronic diseases resulting in increased departmental workloads, staff shortages, and the requirements for increasingly complex diagnostics to support targeted therapies, particularly in oncology [7]. The role of diagnostic AI algorithms has recently been demonstrated by Xu et al. based at a US health network comprising 28 cancer centers [15]. A whole-slide pathology foundation model, Prov-GigaPath, was trained using 171,189 tissue whole slides from more than 30,000 patients that included 31 types of tissue with 1.3 billion 256×256 histopathology image tiles [15]. The study evaluated using a digital pathology benchmark of nine cancer subtyping tasks and 17 diagnostic tasks using clinical data and data from the Cancer Genome Atlas (TCGA) [15]. The model also included vision and language training by incorporating the pathology reports to provide high performance for several digital pathology tasks, showing the potential for diagnostic reporting for real-world data and whole-slide modeling [15].

In September 2024, Wang and colleagues at Harvard Medical School published the findings from a recently designed and versatile ChatGPT-like AI model, the Clinical Histopathology Imaging Evaluation Foundation (CHIEF) system, to undertake multiple diagnostic tasks across multiple types of tumors [16]. The platform used image analysis and is the first to predict and validate patient outcomes across several international patient groups [16,17]. The CHIEF system trained on 15 million images from selected areas of the slides and then trained further on 60,000 whole-slide images of lung, breast, prostate, stomach, colorectal, esophageal, kidney, liver, thyroid, pancreas, cervix, uterus, ovary, testicular, skin, soft tissue, adrenal gland, brain, and bladder tissue samples [16,17]. CHIEF’s performance was tested on 19,400 whole-slide images from 32 independent datasets collected from 24 international hospitals [16]. The investigators reported that CHIEF outperformed other AI methods by >30% on several tasks, which included cancer cell detection, identification of tumor origin, patient outcome prediction, and the identification of molecular targets associated with response to treatment [16]. Other significant findings from this latest study were that, because of the versatile training methods used, CHIEF performed equally well in biopsy material and surgical excision specimens, regardless of the technique used to digitize the tissue section images [16].

Several factors are currently driving the development and use of digital pathology supported by AI analysis. The increased life expectancy in developed countries is associated with an increasing prevalence of chronic diseases, increasing hospital admissions for treatment, and growing demands for pathology testing [18]. Continuing developments in new diagnostic tests to support personalized medicine, the selection of targeted therapies, mainly in oncology, and the use of approved diagnostics are increasing the workload of pathology departments [18]. During the COVID-19 pandemic, digitized diagnostic pathology systems became increasingly used, particularly in the US, where contact between patients and clinicians and between clinicians was restricted [19]. In April 2020, the US FDA enforced a policy for digital pathology devices during the pandemic, increasing their availability for remote slide review and reporting scanned digital images [20]. By 2021, the patient volume returned to pre-pandemic levels, which set the scene for adopting AI-based digital tools and high demand for workflow management [21]. Also, by 2021, there was an increase in regulatory approvals for AI-integrated medical devices, including for pathology diagnosis [19,22].

The increasing prevalence of chronic diseases and the growing range of digital diagnostics applications are rapidly driving demand for the technology and hardware to support digital pathology, which is expected to lead the medical diagnostics market [22]. As of September 2024, an analysis by Fortune Business Insights identified that the global digital pathology market in 2023 was USD 0.37 billion, a projection for 2032 of USD 3.86 billion, and a compound annual growth rate (CAGR) of 16.4% for 2024–2032 [22].

Conclusions

In diagnostic histopathology, AI or machine learning techniques can improve medical diagnostics by improving diagnostic accuracy. The market for technology to support AI in image analysis is rapidly increasing. Recent studies on deep learning algorithms, particularly those designed for tumor diagnostics, have recently shown significant potential and can achieve high tumor detection and grading standards. Also, an important role of AI in diagnostics is the speed and ease of diagnosis when clinical workloads are growing and diagnostic departments are experiencing staff shortages.