Where the Science Currently Stands on AI and Oral Cancer Diagnosis
Introduction
Artificial intelligence is increasingly being presented as part of the future of cancer diagnosis. In oral cancer, there are good reasons for this optimism. Oral cancer remains a significant global health problem, and one of the persistent reasons for poor outcomes is that too many cancers are still diagnosed late. By the time a lesion becomes unmistakably malignant, the disease may already have invaded deeper tissues or spread to lymph nodes.
Earlier diagnosis can therefore make an enormous difference, and AI offers a potentially powerful tool for narrowing this diagnostic gap. But there is still an important distinction between what AI can already do well and what we sometimes imagine it can do.
What AI can already see
Most AI systems being developed for oral cancer diagnosis are fundamentally image-recognition systems. They are trained to examine clinical photographs of oral lesions, microscopic images of biopsied tissue, cytology specimens, or images produced by specialized imaging equipment.
And within this domain, the results can be impressive. Deep-learning systems have demonstrated high sensitivity and specificity for distinguishing oral cancers and potentially malignant lesions from normal tissue. Some experimental systems have reported diagnostic accuracy above 90%, while AI-assisted analysis of digitized histopathology slides has emerged as one of the strongest applications of the technology.
There is also considerable promise in using AI for community screening. Say, for example, a health worker in a community with limited access to oral medicine specialists takes a photograph of a suspicious lesion with a smartphone and receives an automated assessment indicating whether the patient needs urgent referral. In countries and communities where specialists are scarce and patients frequently present with advanced disease, that could be genuinely useful. But there are still important problems on the horizon.
AI is very good at seeing what we can already see
The major weakness of most current systems is also, interestingly, a reason for their greatest strength: they depend heavily on morphology. If a cancer produces a visible ulcer, changes the color or texture of the mucosa, or creates microscopic abnormalities that can be captured on a histopathology slide, an appropriately trained AI system has something to recognize.
But cancer biology does not begin when a lesion becomes visually obvious, and some oral cancers infiltrate beneath the surface while producing relatively little visible mucosal change. These submucosal or early infiltrating lesions present a fundamental problem for image-based AI, resulting in a case where there may simply not be enough information in the photograph for the algorithm to recognize.
No matter how sophisticated an image classifier becomes, it cannot reliably detect biological information that is absent from the image. A similar problem appears in the detection of high-grade epithelial dysplasia, an important precursor to invasive cancer. AI-assisted cytology has shown promise, but detecting these lesions remains difficult partly because abnormal cells in deeper epithelial layers may not be adequately captured by superficial sampling.
There is another problem: many AI systems are developed using datasets from a single hospital or research centre. A model that performs extremely well on the population it was trained on may perform differently when exposed to patients, cameras, laboratory techniques, or clinical environments elsewhere. So, while the numbers reported in individual studies can be striking, translating them into reliable real-world diagnostic systems remains considerably more complicated.
The next leap will require more than better photographs
The future of AI-assisted cancer diagnosis must shift focus from being just about teaching computers to become progressively better observers of visible disease. We also need to give them access to the biology happening before disease becomes visible.
Cancer begins as a biological process long before it becomes an obvious photograph: genes change their patterns of expression, proteins change, epigenetic programs are altered, cellular metabolism is rewired, immune cells begin interacting differently with abnormal cells, and the surrounding tissue and extracellular matrix change as a tumour develops and prepares to invade. Increasingly, we can (and we must) measure these processes alongside.
Instead of an AI system receiving only an image of a suspicious lesion, future diagnostic systems could integrate that image with molecular information from the same patient: gene-expression patterns, proteins, epigenetic changes, metabolic signatures, or other biomarkers. This is the promise of multiomics.
At the same time, technologies such as single-cell RNA sequencing and spatial transcriptomics are giving researchers increasingly detailed maps of the tumour microenvironment. They allow us to examine whether cancer cells are present, what kinds of cells surround them, how immune cells are behaving, how fibroblasts are changing, and how the architecture of the tissue is being reorganised.
Molecular imaging offers another possibility. Rather than waiting for cancer to alter the physical appearance of tissue, imaging technologies can potentially detect functional or molecular changes occurring beneath the surface. Combined with AI, this could eventually help address one of the most important blind spots of conventional image-based systems: disease that is biologically present but visually inconspicuous.
Boluwatife OLU Afolabi
This article is a shorter adaptation of our preprint, “Artificial Intelligence and Oral Cancer Diagnosis: Current Evidence, Critical Limitations, and the Basic Science Priorities Essential for Improving Diagnostic Accuracy.” The full scientific paper, including the underlying literature and references, is available as a preprint here: Preprints.org — full paper