How Visual AI Is Transforming Cancer Detection and Blood Cell Diagnosis

Md Muzahidul Islam

A radiologist reaches the end of a long day after examining hundreds of mammograms. Most are little more than shades of gray. Yet in a few, there is a faint shadow no larger than a grain of rice. Detecting it can save a life, missing it can cost one. For decades, the only safeguard against human fatigue was to have another equally exhausted clinician review the same image. This is the silent challenge visual AI was designed to address and evidence emerging from clinical practice suggests it is beginning to deliver.

What once seemed like a promise is now being backed by hard evidence from real-world trials. Researchers in Sweden conducted the world’s first large randomized trial of AI-assisted breast cancer screening. Using AI software to interpret mammograms, the study detected 29% more cancers without increasing false-positive results, while reducing radiologists’ reading workload by half. These findings were not drawn from a small, controlled dataset but from more than 80,000 women participating in the country’s routine public screening programme.

Similar results have emerged from the United Kingdom. A study evaluating deep-learning software across more than 26,000 mammography screening cases found that the AI system performed on par with a highly experienced radiologist. Moreover, when paired with a single radiologist, the software matched the diagnostic accuracy of the traditional two-radiologist reading model. Rather than replacing medical expertise, AI effectively enabled one specialist to perform the work of two.

A second reader that never tires fundamentally changes the economics and efficiency of cancer screening. In oncology, early detection is everything. Tumours identified at an early stage are often treatable before they spread, while cancers missed during initial screening and discovered later tend to be more aggressive. At the same time, healthcare systems worldwide face a shortage of radiologists even as imaging volumes continue to rise. AI’s ability to prioritize suspicious scans allows specialists to focus their attention where it is needed most.

Regulators have taken notice. Medical imaging has become the fastest-growing area for regulatory approval of medical AI, with the U.S. Food and Drug Administration (FDA) having cleared approximately 1,000 AI-enabled medical devices, the majority of which are designed for image interpretation.

Another field where visual AI is advancing rapidly is blood analysis. Traditionally, blood smear examination requires a laboratory specialist to manually inspect a drop of blood under a microscope and count different types of blood cells. The process is time-consuming, and even experienced professionals may reach different conclusions when examining the same slide. AI systems trained on microscopic cell images have demonstrated the ability to classify white blood cells—including neutrophils, lymphocytes and other cell types—with accuracy exceeding 99% in a matter of seconds.

Such speed can be critical when diagnosing conditions such as leukemia or serious infections. One of the studies underpinning this article describes a lightweight neural network capable of running on modest computing hardware rather than requiring large-scale servers. This makes it possible for the technology to operate close to the patient’s bedside, where blood samples are collected.

Yet the greatest challenge has never been accuracy—it has been trust. Doctors must understand and, when necessary, challenge an AI system’s conclusions before acting on them. A “black box” that simply labels an image as cancerous without explaining its reasoning is difficult to defend before patients or in clinical practice.

To address this, researchers have developed visualization techniques that overlay colour-coded heat maps onto medical images, highlighting the specific regions that influenced the AI’s decision. Recent studies combine these visual explanations with AI classifiers, enabling clinicians to determine whether the model focused on the relevant cell or merely on an irrelevant artifact or smudge.

The white blood cell study referenced earlier employs Grad-CAM and Grad-CAM++, methods that transform the model’s internal attention into visual evidence that clinicians can inspect and verify. The principle is straightforward: an AI system that explains its reasoning can be evaluated, improved and trusted; one that cannot remains difficult to validate.

Another persistent challenge is poor-quality data and the solution is less glamorous than the headlines surrounding AI. Medical datasets are inherently imbalanced, typically containing far more healthy samples than diseased ones. Left unchecked, AI models often learn the simplest—but least useful—strategy: predicting that every case is healthy while still appearing highly accurate because normal cases vastly outnumber abnormal ones.

One widely adopted solution is ADASYN, a technique that generates synthetic examples of underrepresented disease cases until the dataset becomes sufficiently balanced for effective learning. A companion study on breast cancer prediction examined how such balancing techniques influence the performance of multiple machine-learning approaches, including distance-based models, decision trees, ensemble methods and boosting algorithms.

The findings carry important clinical implications. Both the choice of AI model and the method used to balance training data can significantly influence outcomes—potentially determining whether a patient’s cancer is detected or overlooked.

None of this transfers medical decision-making from physicians to machines. The promise of visual AI is far more practical. It equips clinicians with a faster, sharper, and more transparent second opinion—one whose reasoning can be examined rather than blindly accepted. Just as the microscope did not replace the pathologist, visual AI will not replace the physician. What it changes is the speed with which patients receive answers and the likelihood that a tiny, life-threatening shadow is detected while there is still time to act.

Md Muzahidul Islam is a researcher in Information Systems at Lamar University.