Most AI systems for tumor detection are trained on examples of both normal and abnormal scans. This approach requires large labeled datasets of tumor images, which are expensive and time-consuming to collect and annotate. This paper takes a completely different approach: the system learns only from normal breast scans and then detects tumors as anomalies.
The method, called PICARD (Pluralistic Image Completion for Anomaly Recognition and Detection), works by learning what a normal breast scan looks like. At test time, if a region of the scan cannot be reconstructed to look normal, the system flags it as a likely tumor.
PICARD was designed for Digital Breast Tomosynthesis (DBT) - a 3D mammography technology that takes images from multiple angles to create a detailed 3D picture of breast tissue, reducing the problem of overlapping tissues that can obscure tumors in traditional 2D mammograms.
Supervised AI systems for cancer detection require radiologists to carefully label every tumor in every training scan - a process that can take hours per case and requires specialized expertise. This creates a fundamental bottleneck: good AI needs lots of data, but creating that data is expensive.
Unsupervised anomaly detection sidesteps this bottleneck. Only normal scans are needed for training - and there are typically far more normal scans available in clinical databases than scans with confirmed tumors.
Digital Breast Tomosynthesis generates high-resolution 3D volumes, making tumor localization more precise than 2D mammography but also creating a computational challenge: the images are large, require significant processing, and the tumors can appear subtle against complex normal breast architecture.
PICARD is built on a Generative Adversarial Network (GAN) trained to perform image completion - given a breast scan with a region masked out, the network fills in the masked region to look like plausible normal tissue. After training on only normal scans, the network becomes an expert at generating realistic normal breast tissue.
At test time, every region of a patient's scan is systematically masked and then reconstructed. The key insight is the Minimum Completion Distance (MCD) metric: if the real tissue in a region looks very different from all the reconstructed 'normal' versions, that region is likely abnormal. The MCD measures this discrepancy.
A crucial technical innovation is using spatial dropout during reconstruction to generate multiple diverse completions ('pluralistic completion') for each masked region. This means MCD measures the best possible normal reconstruction - if even the best normal guess is far from the real tissue, that is strong evidence of abnormality.
Standard image completion models produce a single reconstruction. But normal breast tissue is highly variable - many different tissue patterns are all 'normal'. If the model only generates one reconstruction, it might miss the specific normal pattern that best matches the real tissue, creating false positives.
Spatial dropout introduces randomness into the completion process, causing the GAN to generate different but all equally plausible normal reconstructions each time. By taking the minimum distance across all these reconstructions (MCD), PICARD uses the most favorable comparison for each region.
This reduces false positives: normal tissue that simply looks unusual is more likely to have at least one reconstruction that closely matches it. Only truly abnormal tissue - which no normal reconstruction can match - receives a high anomaly score.
Evaluated on pixel-level tumor localization in DBT scans, PICARD achieved an AUC of 0.875. This means the anomaly maps generated by PICARD correctly identified tumor locations with 87.5% discrimination ability.
This outperformed all competing unsupervised anomaly detection methods by more than 10 percentage points in AUC. The closest competitor used a single-completion approach (without spatial dropout), demonstrating that pluralistic completion is the key differentiating factor.
The performance is particularly impressive given that PICARD was never shown a single tumor image during training. It effectively learned the definition of 'abnormal' by exclusion - anything that cannot be made to look normal is abnormal.
An unsupervised system like PICARD could be deployed in new clinical environments without needing to collect local annotated tumor datasets. A hospital could install the system, train it on their existing archive of normal screenings, and immediately begin using it for anomaly flagging.
This is particularly valuable for DBT screening programs, which generate very high volumes of 3D images. AI systems that can pre-screen studies and prioritize suspicious cases for radiologist review could significantly reduce reading workload.
Because PICARD produces a pixel-level anomaly map (not just a scan-level flag), it can highlight the specific location of suspicious regions for the radiologist, serving as a precise visual guide.
PICARD demonstrates that unsupervised anomaly detection can achieve radiologically meaningful performance for breast tumor detection without any labeled abnormal examples. The pluralistic completion approach is the key technical advance that makes this possible.
The approach is generalizable - the same framework could potentially be applied to other types of medical scans and other cancer types where normal tissue patterns are learnable and abnormalities manifest as deviations from normal.
Future directions include combining PICARD's anomaly maps with downstream supervised classifiers to further improve precision, and evaluating performance across diverse patient populations and DBT scanner models.