Cancer diagnostics generates enormous volumes of data every day: imaging scans, pathology slides, genomic assays, and clinical records. In principle, this data could train highly capable machine learning classifiers to assist with cancer detection, grading, and outcome prediction. In practice, the bottleneck is not the volume of data but the availability of labeled data, where medical experts have annotated each sample with the clinically correct classification.
The dominant approach in medical AI is supervised learning (SL), which requires a fully labeled dataset for training. Obtaining labels requires experienced specialists such as radiologists and pathologists to review each image or sample and assign a diagnosis. This process is time-consuming, expensive, and increasingly constrained as physician workloads grow and the number of cancer patients rises in an aging population.
The fundamental mismatch in medical AI is that unlabeled data is abundant and growing while labeled data is scarce and expensive to produce. For example, a hospital may acquire thousands of prostate MRI scans annually, but only a fraction will receive the detailed expert annotation needed to train a supervised classifier. The challenge of closing this gap has prompted interest in approaches that can learn from both labeled and unlabeled data.
Semi-supervised learning (SSL) addresses this mismatch by training models on a small labeled dataset augmented by a large unlabeled dataset. Rather than discarding the information contained in unlabeled samples, SSL algorithms use them to learn the underlying structure of the data, potentially achieving classification performance comparable to fully supervised approaches while requiring only a fraction of the labeling effort.
SSL occupies the conceptual space between supervised learning, which uses only labeled data, and unsupervised learning, which uses only unlabeled data to find hidden structure. SSL combines both: labeled data trains the initial classifier while unlabeled data from the same distribution is incorporated to constrain and improve the decision boundary that separates classes.
SSL rests on three interconnected assumptions. The smoothness assumption states that data points with similar feature vectors should receive similar class labels. The manifold assumption holds that high-dimensional data points cluster along lower-dimensional structures, and points on the same manifold belong to the same class. Together, these imply that the low-density assumption holds: optimal decision boundaries should pass through regions where data points are sparse rather than through dense clusters, and unlabeled data can help identify those sparse regions.
When these assumptions are met, adding unlabeled data from the same distribution as the labeled data tightens the decision boundary in a way that improves generalization. The key practical requirement is distribution matching: the unlabeled data must come from the same population as the labeled data. For example, an SSL model for prostate cancer Gleason grading should use unlabeled prostate biopsy slides with the same staining protocol, not slides from a different tissue type or preparation method.
When distribution assumptions are violated, SSL can degrade performance rather than improve it, because the mismatched unlabeled data introduces noise that confuses rather than helps the classifier. This means SSL is not a universally applicable improvement over supervised learning; careful design and validation are required. SSL methods are broadly divided into inductive methods, which generalize to new unseen data, and transductive methods, which only make predictions within the training sample. Inductive methods are preferred for clinical applications where the model must work on new patients.
Histopathology is a natural domain for SSL because whole-slide images (WSI) of tumor tissue are routinely acquired in abundance but expert annotation of individual image regions is labor-intensive. SSL studies in histopathology address both patch-level classification, identifying cancerous regions within a slide, and slide-level classification, determining whether a sample as a whole is malignant.
A mean teacher model trained on over 13,000 colorectal cancer WSI achieved AUC comparable to a multi-pathologist benchmark, with substantial SSL improvement over supervised learning when labels were limited. Critically, the performance gap between SSL and SL disappeared when a fully labeled dataset was available, confirming that SSL's advantage is specifically in the label-scarce regime. A student-teacher chain model for the same task achieved comparable performance using only 0.5% labeled data.
For prostate cancer Gleason scoring, two SSL approaches are highlighted. A teacher-student chain with pseudo-labeling reported performance improvements over a student-only baseline. A pixel-based approach using expectation maximization with a fully convolutional encoder-decoder network incorporating both internally annotated and externally weakly annotated images also achieved improvements over models trained on a fully labeled dataset alone. Both confirm that SSL can reduce the annotation burden for Gleason grading without sacrificing accuracy.
Other SSL histopathology studies achieved a 19% accuracy improvement for bladder cancer detection and grading using only 3% labeled data with consistency regularization and self-ensembling, similar accuracy to supervised approaches in breast cancer classification using half the number of labels, and high-performance PD-L1 scoring in non-small cell lung cancer using an auxiliary classifier GAN. These results collectively show SSL's broad applicability across cancer types and histopathology tasks.
In radiology, SSL has been extensively applied to lung nodule detection and classification on CT scans. Multi-tasking CNNs that simultaneously learn nodule segmentation and false positive reduction, semi-supervised adversarial autoencoders for benign versus malignant nodule classification, and transfer learning with semi-supervised feature matching all report high accuracy on chest CT benchmarks. A common caveat is that performance on independent validation sets is lower than on internal datasets, underscoring generalization challenges.
For breast cancer detection, co-training approaches on mammograms and a joint weakly and strongly supervised self-training framework for ultrasound imaging demonstrate SSL's value across modalities. One ultrasound study achieved comparable accuracy using only ten strongly annotated images combined with many weakly annotated ones, versus 800 strongly annotated images in the supervised baseline, representing a substantial reduction in labeling burden.
In radiotherapy planning, SSL has been applied to identifying target volumes for postoperative tumor bed irradiation in breast cancer, and to detecting residual tumor in postoperative brain MRI for high-grade glioma patients, where semi-supervised decision forests improved both performance and computation time compared to conventional segmentation.
For non-image genomic data, SSL has been applied to survival prediction from gene expression data across breast, lung, gastric, and liver cancers, cancer recurrence prediction using low-density separation, and tumor subtype identification from multi-omics data. One affinity fusion network approach for classifying primary tumor sites achieved high accuracy with training on less than 1% labeled data, demonstrating SSL's potential to extract diagnostic signal from genomic data without extensive expert annotation.
The most consistent finding across reviewed SSL studies is that performance advantages over supervised learning diminish as labeled data increases. SSL provides the greatest benefit specifically in label-scarce settings; when fully labeled datasets are available, the gap between SSL and SL shrinks or disappears. This means that SSL should be evaluated relative to a well-tuned supervised baseline, not just on absolute accuracy, and that the value of SSL depends on how severe the labeling constraint is for a given clinical problem.
A major methodological concern is the frequent absence of a fair comparison between SSL and SL baselines in published studies. Many papers report SSL performance without providing a supervised learning baseline trained on the same labeled data, making it impossible to assess the actual benefit of adding unlabeled samples. Fewer still report equivalent hyperparameter tuning for both approaches or vary the amount of labeled and unlabeled data systematically to characterize the SSL-SL performance trade-off.
Distribution mismatch between labeled and unlabeled data is the primary technical failure mode. Differences in imaging equipment, staining protocols, image resolution, or patient population between the labeled and unlabeled subsets can cause SSL to degrade performance rather than improve it. For multicenter applications, where images from different institutions are pooled, careful preprocessing and domain adaptation strategies are needed to ensure that unlabeled data from different sites does not introduce confounding noise.
The explainability problem is particularly acute in SSL. Supervised learning models are already difficult to interpret, but SSL adds a further layer of opacity because information is also derived from unlabeled samples whose class assignments are inferred rather than expert-provided. Clinical specialists in radiology and pathology are unlikely to trust and adopt models whose decision-making process they cannot trace, and regulatory and ethical frameworks increasingly require AI tools to provide interpretable outputs before they can enter routine clinical practice.
The review identifies three priority areas for advancing SSL in cancer diagnostics. First, data sharing is essential. Most existing SSL research focuses on high-prevalence cancers such as breast, lung, and colorectal cancer, where single institutions can accumulate enough data for experiments. Expanding to rare tumor entities and hematological malignancies requires consortium-level data sharing of both labeled and unlabeled samples across institutions.
Second, the field needs to move from single-center retrospective studies to multicenter prospective validation. SSL models trained on heterogeneous data from multiple institutions and scanner types are more likely to generalize to new clinical environments. Crowd-sourcing model development across research groups would simultaneously expand the unlabeled data pool and accelerate validation in diverse patient populations.
Third, explainability mechanisms must be built into SSL models rather than added as an afterthought. Specific needs include making intermediate steps such as pseudo-label assignment on unlabeled samples traceable, providing per-prediction confidence estimates, and generating visual explanations comparable to how radiologists or pathologists describe their reasoning. Without these, SSL models will face substantial barriers to regulatory approval and clinical acceptance.
Despite these challenges, SSL represents one of the most practical near-term approaches to overcoming the labeled data bottleneck that limits cancer AI. The clinical data already exists in electronic health records, PACS systems, and genomic databases; the question is whether SSL can unlock the diagnostic signal contained in the unlabeled majority of that data. With rigorous study design, multi-institutional collaboration, and attention to explainability, SSL has the potential to substantially expand what is achievable with current volumes of annotated medical data.