Researcher unveils AI tool for detecting 26 cancers with near-perfect accuracy

A new study published in Diagnostics has unveiled a browser-based artificial intelligence framework capable of identifying 26 different types of cancer with near-perfect accuracy, in what researchers describe as a breakthrough that could expand access to advanced cancer diagnostics globally.

The study, led by researcher Blessing Ndubuisi, introduced what was described as a first-of-its-kind system that enables high-performance diagnostic analysis through a standard web browser, potentially removing long-standing barriers associated with expensive hardware and complex computing infrastructure.

According to the study, the innovation addresses a major challenge in deploying artificial intelligence in healthcare, particularly in low-resource settings where advanced diagnostic tools are often unavailable.

Unlike conventional deep learning models that depend on powerful servers and costly computing equipment, the framework allows analysis to take place directly on a user’s local device, including computers and tablets, without transmitting data externally.

The research, which relied on more than 130,000 medical images, involved the development and fine-tuning of a neural network trained to detect subtle patterns associated with multiple cancers.

Findings from the study showed the model achieved a Top-1 accuracy of 99.85 per cent, indicating that its first prediction was correct in nearly all cases evaluated.

It also recorded a Top-5 accuracy of 100 per cent, meaning the correct diagnosis consistently appeared among the model’s top recommendations.

The study further noted that the framework outperformed several conventional AI models while requiring significantly lower computing power and memory.

Ndubuisi said the study also prioritised transparency in diagnosis by integrating explainable artificial intelligence features to improve clinical confidence in the system’s predictions.

Using a technique known as Gradient-weighted Class Activation Mapping, also called Grad-CAM, the framework produces heatmaps that highlight the specific areas of a medical image that influenced a diagnosis, including clusters of abnormal cells.

According to the study, the visualisations showed that the model’s predictions were guided by histopathologically relevant regions, providing a transparent rationale that could help clinicians cross-check and trust the system’s recommendations.

A major highlight of the innovation is its use of edge computing through TensorFlow.js, which enables the entire diagnostic process to occur inside a web browser without sending patient images to remote servers.

The researchers noted that this approach offers multiple advantages, including elimination of infrastructure costs tied to specialised hardware, stronger patient privacy protections since data remains on local devices, and broader accessibility for clinicians regardless of location.

The study said the framework could help bridge inequities in cancer detection, particularly in underserved regions, by making advanced diagnostic tools available through widely accessible digital platforms.

It added that beyond its immediate diagnostic potential, the framework establishes a scientific benchmark for future global cancer screening tools and demonstrates a practical pathway toward equitable distribution of life-saving medical technologies.