Perhaps no ambition has loomed larger in healthcare than the cure for cancer. In 2025, cancer is expected to claim 618,000 lives in the United States alone, and by 2050, annual new cases worldwide are projected to exceed a staggering 35 million. For decades, our efforts have centered around developing new therapeutics, yet progress remains painfully slow, prohibitively expensive, and often benefits only a small subset of patients. 

That reality has demanded a paradigm shift in the fight against cancer: catch it before a cure is even required. Powered by AI, recent developments in cancer diagnostics have shown remarkable progress, with the potential to drive mortality downward by detecting tumors in their early stages.

The numbers speak for themselves. Consider lung cancer: if detected in its final stage—which most cases are—the survival rate is 40%, but if found in its first stage, the survival rate is 99%. In colorectal cancer, five-year survival rates jump from 18.4% to 92.3% when diagnosed early. This pattern persists across nearly all cancer types, underscoring that simply detecting cancer in its earliest stages can rival, and surpass even, some of our most advanced therapeutics.

Currently, there are only five widely accepted cancer diagnostic tests, each aimed at a specific type of cancer: breast, lung, prostate, colorectal, and cervical. While these tests have saved countless lives, they struggle with limitations, including low sensitivity (missed cancers), low specificity (false alarms), procedural invasiveness, and limited accessibility. Most sobering of all, an estimated 70% of cancer deaths arise from cancers without existing screening tests. 

That is where multi-cancer early detection (MCED) tests come in. Rather than relying on costly scans or invasive biopsies, they begin with a simple blood draw. Next, researchers collect and analyze molecular traces shed by tumors in the blood (DNA, proteins, metabolites, etc.). AI models then use this data to form predictions about the presence of cancer, and, in some cases, pinpoint the tissue of origin. In addition to their improved accessibility and cost, MCED tests have the potential to detect a far wider range of cancers than existing diagnostic tests.

One especially promising cancer early detection test has been developed by DELFI diagnostics. Their approach examines cell-free DNA (cfDNA)—fragments of genetic material released into the bloodstream as cells die. Within that pool lies circulating tumor DNA (ctDNA), the fraction of cfDNA originating from cancer cells. DELFI’s key insight is that the fragment lengths of ctDNA are measurably shorter and follow distinct fragmentation patterns compared with cfDNA from healthy cells—a disparity which their model uses to make predictions. 

After taking a blood draw and extracting the cfDNA, DELFI performs low-coverage whole-genome sequencing to capture millions of DNA reads at minimal cost. From those reads, the pipeline generates a “fragmentation profile,” with hundreds of features useful for differentiating between cancer and healthy cells. Finally, DELFI feeds these fragmentomics data, along with additional DNA-related features like copy number changes, into a machine-learning classifier that predicts whether cancer is present. 

DELFI’s first commercial assay, FirstLook Lung, provides a non-invasive and easily accessible alternative to existing low-dose CT (LDCT) scans for lung cancer. The classifier was trained and clinically validated through the prospective case-control DELFI-L101 study, which included 958 participants across 49 sites—a sample broadly matching the lung cancer screening-eligible population per United States Preventive Services Task Force (USPSTF) guidelines. In an independent validation cohort, FirstLook Lung achieved a sensitivity of 80% and a specificity of 58%.

Those metrics suggest that FirstLook Lung would be an effective prescreen for LDCT imaging. Although 58% specificity (42% false positives) may seem high, every study participant already met the lung cancer screening criteria; that is, the 42% flagged by FirstLook would have undergone imaging anyway. Crucially, the assay’s negative predictive value is 99.8%, meaning that only a fraction of a percent of people with a negative test result actually harbor LDCT-detectable lung cancer. 

Significant hurdles remain. On the regulatory side, no MCED tests have been approved by the FDA to date. Instead, some assays—like DELFI’s FirstLook—are offered as laboratory-developed tests under the Clinical Laboratory Improvement Act (CLIA) regulations, which allow doctors to order them for clinical use. Furthermore, on a biological level, detecting early-stage cancers that shed extremely low concentrations of ctDNA can be intrinsically difficult, pushing the limit of detection. And even after detecting the presence of cancer, the task of pinpointing its tissue of origin is an ongoing challenge. 

Still, the promise of these AI-powered cancer screening tests cannot be ignored. As larger genomic datasets across different types of cancers become available and algorithms continue to mature, test accuracy will only climb. Combining different biomarkers is already boosting performance. For populations without easy access to advanced imaging and invasive biopsies, a simple blood draw offers a scalable alternative. And at the same time, healthcare systems can reduce their economic burden by avoiding costly, unnecessary late-stage interventions. The future for cancer detection isn’t just bright—it heralds a fundamental shift in our approach to the fight against cancer itself.