What Is Tomosynthesis with CAD in Breast Imaging?

Tomosynthesis with CAD combines a three-dimensional mammography technique called digital breast tomosynthesis (DBT) with computer-aided detection or diagnosis software that flags suspicious areas in the images. The goal is straightforward: give radiologists a clearer, layered view of breast tissue and then use software to highlight spots that might be cancer, all in the same sitting. The combination has become one of the most actively researched areas in breast imaging over the past decade, and recent advances in artificial intelligence have changed what the “CAD” part of the equation can actually do.

How Tomosynthesis Creates a Layered Image

A standard digital mammogram compresses the breast and takes a single flat X-ray. Tomosynthesis works differently. The X-ray tube moves in an arc over a limited angular range, collecting multiple low-dose projection images of the compressed breast. Those projections are then reconstructed into a stack of thin slices, giving radiologists the ability to scroll through breast tissue layer by layer rather than looking at everything squished into one flat picture.¹ That third dimension is the key advantage: overlapping tissue that might hide a small mass on a flat mammogram can be separated out on tomosynthesis, making subtle findings easier to spot.²

The trade-off is data volume. A single tomosynthesis exam produces dozens of image slices per breast. Reviewing all of those slices takes longer than glancing at a two-view mammogram, and the sheer number of images creates a natural opening for software that can pre-scan the stack and point the radiologist toward areas that deserve extra attention. That is where CAD enters the picture.

What Computer-Aided Detection Actually Does

CAD software analyzes the tomosynthesis images and places marks, sometimes called prompts, on areas it considers suspicious. Those marks are not diagnoses. They are suggestions: “look here more carefully.” The radiologist still makes the final call about whether to recall a patient for further workup. In its earliest form, CAD for standard mammography used relatively simple pattern-recognition algorithms to flag shapes that resembled masses or clusters of tiny calcium deposits. The problem was that conventional CAD programs were noisy. They flagged a lot of areas that turned out to be nothing, and across years of use they did not clearly improve diagnostic accuracy for standard mammograms.³

The newer generation of CAD tools built for tomosynthesis takes a fundamentally different approach. Instead of handcrafted rules about what a lesion looks like, these systems use deep learning, a type of artificial intelligence that learns patterns directly from large sets of labeled images. The shift from rule-based CAD to deep-learning CAD has been transformative: the software now produces fewer false marks per image and generates a numerical score reflecting how suspicious each finding is.³ As of mid-2024, six AI-based CAD tools for tomosynthesis had received FDA clearance, each designed to detect suspicious lesions and assign malignancy likelihood scores at both the lesion and exam level.⁴

Does It Actually Improve Cancer Detection?

The evidence from multi-reader studies generally points in a positive direction, though the gains are often modest. In one multi-reader, multi-case study, radiologists using a concurrent CAD system while reading tomosynthesis exams saw their average sensitivity rise from about 85% to 87%, with no meaningful drop in specificity.⁵ A separate study using a newer AI-based tool found a similar bump: radiologists’ sensitivity went from roughly 85% without AI to about 88% with it, again with no evidence that specificity suffered.⁶

A few percentage points of sensitivity may not sound dramatic, but in a screening population where cancers are uncommon, those points translate into real cases caught earlier. The gains tend to be most noticeable for soft-tissue lesions such as masses rather than for calcification clusters, which radiologists already catch at high rates. One early CAD system tested on tomosynthesis achieved about 95% sensitivity for calcification clusters and 89% for masses, with an average of roughly 2.7 false-positive marks per image view.⁷ That false-positive rate is something developers have been steadily working to reduce, because every unnecessary mark adds to the radiologist’s cognitive load.

Reading Time and Workflow

One of the practical headaches of tomosynthesis is how long it takes to read. Scrolling through dozens of slices per breast, times two breasts, times hundreds of patients a day, adds up fast. CAD tools designed for tomosynthesis address this in two ways. First, they can direct the reader’s eye to specific slices, reducing aimless scrolling. Second, some systems generate a “top slice” or summary view that highlights the software’s findings, so the radiologist can start with the most important areas.

In a multi-reader study, AI support cut the average reading time per tomosynthesis exam from about 41 seconds to about 36 seconds, a reduction of roughly 12%.⁸ That five-second difference sounds trivial until you multiply it across a full day of screening reads. For a radiologist interpreting several hundred exams in a shift, reclaiming five seconds per case frees up meaningful time. The effect is amplified in triage workflows, where AI scores are used to sort cases before a human ever looks at them. Researchers have explored using the AI’s confidence score to filter out clearly normal exams from the worklist entirely, so radiologists spend their time only on cases the algorithm considers uncertain or suspicious.⁹

Recall Rates and False Positives

Being called back after a screening mammogram for additional imaging is stressful, and one of the promises of better technology is reducing unnecessary recalls. The picture here is mixed. Adding CAD to tomosynthesis sometimes introduces a small uptick in recall for non-cancer cases, because the software’s marks can nudge a radiologist toward recommending follow-up they might otherwise have skipped. One pilot study found a slight, non-significant increase in recall for non-cancer cases when CAD was used.¹⁰

On the other hand, AI-based CAD has shown the ability to dramatically lower recall rates in specific clinical contexts. In a study of women undergoing surveillance mammography after breast-conserving therapy for cancer, adding AI-CAD to digital mammography dropped the recall rate for the treated breast from about 11% down to roughly 2%, while overall accuracy went up.¹¹ That is a dramatic improvement in a population where post-surgical changes in the breast make interpretation especially tricky and false alarms are common. However, that same study found that sensitivity for detecting actual recurrences was lower with AI-CAD compared to standard mammography alone, a trade-off that highlights how reducing false positives and catching true cancers can pull in opposite directions.

Synthetic Mammograms and Radiation Dose

When tomosynthesis first arrived, many facilities performed it alongside a conventional 2D mammogram, which roughly doubled the radiation dose per exam. The solution was the synthetic mammogram: a computer-generated flat image reconstructed from the tomosynthesis data, eliminating the need for a separate 2D exposure. AI-based CAD has been tested on these synthetic images, and the results are encouraging. One study found that using AI-CAD with synthetic mammograms plus tomosynthesis performed at least as well as conventional mammography plus tomosynthesis, with a lower radiation dose and shorter reading time.¹² This matters for screening programs where millions of women are imaged every year and even a small per-exam dose reduction adds up across the population.

When the Algorithm Disagrees with Itself

An underappreciated wrinkle arises when CAD software analyzes both the 2D mammogram and the tomosynthesis stack from the same patient and reaches different conclusions on each. Researchers studying this discordance found only moderate agreement between the two modalities, with roughly a third of mass-type lesions and close to 40% of calcification cases producing conflicting scores between the mammogram and the tomosynthesis read.¹³ AI-powered reclassification models were able to resolve a large share of these conflicts, correctly reclassifying over 80% of discordant mass cases. But the existence of inter-modality disagreement is something radiologists using these tools need to be aware of, because a clean bill of health on the 2D view and a flagged lesion on the 3D view (or vice versa) requires careful judgment.

Dense Breast Tissue

Breast density is one of the biggest challenges in mammographic screening. Dense tissue appears white on a mammogram, and so do many cancers, making detection harder. Tomosynthesis has been promoted as especially useful for dense breasts because its layered approach can separate dense tissue from underlying lesions. Research on deep-learning models applied to tomosynthesis has found that error rates remained fairly stable across density categories, though there was a slight decrease in sensitivity in the densest breasts accompanied by a small rise in specificity. In practical terms, the technology helps with dense tissue, but it does not entirely eliminate the masking problem. Women with extremely dense breasts may still benefit from supplemental screening with ultrasound or MRI even when tomosynthesis with CAD is used.

The Automation Bias Problem

Perhaps the most important caution around CAD and AI in breast imaging is not about the technology itself but about how humans interact with it. Automation bias is the tendency of a person to trust a computer’s suggestion even when it is wrong. A study examining this phenomenon in mammography found that radiologists of all experience levels were significantly influenced by incorrect AI suggestions. When the AI correctly categorized a case, readers across experience levels scored above 80% accuracy. When the AI was wrong, accuracy plummeted: inexperienced readers dropped to roughly 20% correct, and even highly experienced readers fell to about 45%.¹⁴

Inexperienced radiologists were especially vulnerable, being significantly more likely to follow an incorrect AI suggestion to upgrade a case’s suspicion level. This is not a quirk of one study. Earlier research on conventional CAD showed a similar pattern: when CAD failed to flag a cancer, readers’ sensitivity dropped compared to reading without CAD at all, as though the absence of a prompt functioned as a false reassurance.¹⁵ The implication is clear: if radiologists lean on the software too heavily, a missed prompt can become a missed cancer. Training programs increasingly emphasize that CAD should be treated as a second opinion, not a safety net.

Algorithmic Bias Across Patient Populations

AI tools are only as fair as the data they were trained on, and breast imaging AI is no exception. A study of a commercially available AI algorithm used on negative screening tomosynthesis exams found that false-positive case scores were about 50% more likely in Black patients compared to White patients, and less likely in Asian patients. Older patients (ages 71 to 80) were roughly twice as likely to receive a false-positive score as those aged 51 to 60. And patients with extremely dense breasts were nearly three times as likely to receive a false-positive risk score as those with fatty breasts.¹⁶

These disparities matter because a higher false-positive rate in certain populations means more unnecessary callbacks, more biopsies, more anxiety, and more cost for those patients. If left unaddressed, AI tools risk reinforcing existing inequities in breast cancer screening. Developers and regulatory bodies are increasingly expected to report algorithm performance broken down by race, age, and breast density, but at the moment there is no universal requirement to demonstrate equitable performance before an AI tool reaches the clinic.

Newer Algorithms That Learn from Prior Exams

One of the most promising directions in CAD research is incorporating temporal information, meaning the algorithm compares a patient’s current tomosynthesis exam to her previous ones and looks for changes over time. Radiologists have always done this manually, pulling up prior mammograms side by side. But doing it computationally across 3D image stacks is a different challenge. Researchers have developed deep-learning frameworks that explicitly integrate temporal change into the detection model, with the goal of distinguishing a new or growing lesion from a stable finding that has been present for years.¹⁷ This approach has the potential to reduce false positives from stable benign findings while improving detection of interval changes that might signal early cancer.

What Patients Think About AI Reading Their Mammograms

A pilot survey found that most patients are not opposed to AI being part of their mammogram interpretation, but comfort levels are closely tied to understanding. About 64% of surveyed patients said they would feel more comfortable with AI if they better understood how it was being used.¹⁸ This suggests that imaging centers rolling out AI-assisted tomosynthesis would benefit from proactively explaining the technology to patients rather than letting them learn about it after the fact. The key messaging point is that AI serves as an extra set of eyes for the radiologist, not a replacement for human judgment. Given the automation bias findings described above, that framing is not just good communication but also an accurate description of how the tools should be used.

How the Exam Differs from a Standard Mammogram

From your perspective as a patient, a tomosynthesis exam feels almost identical to a regular mammogram. Your breast is still compressed between two plates, and the whole thing takes only a few seconds longer per view because the X-ray tube needs to sweep its arc. You will not know whether CAD software is analyzing your images afterward. The results come back the same way: a letter or a phone call telling you the exam was normal or that additional imaging is needed. The behind-the-scenes difference is that your radiologist reviewed a three-dimensional image stack with AI annotations rather than a flat picture alone.

If your facility uses synthetic mammograms generated from the tomosynthesis data, the overall radiation dose is comparable to a standard mammogram. If they acquire both a conventional 2D mammogram and the tomosynthesis sweep, the dose is somewhat higher, though still within safety limits set by regulatory agencies. You can ask the technologist which protocol your facility uses if dose is a concern.

Insurance coverage for tomosynthesis has expanded significantly over the past several years. Most private insurers and Medicare now cover it for screening. Some states mandate coverage by law. CAD analysis is generally bundled into the facility’s interpretation and is not billed separately to you, though billing practices vary. If you are unsure, your imaging center’s scheduling staff can confirm before your appointment.