Deep learning has revolutionized medical imaging, but its progress is tethered to a grueling requirement: thousands of pixel-perfect, expert-annotated masks. In clinical practice, this "expert bottleneck" is often insurmountable. Radiologists’ time is expensive, and rare pathologies offer few examples to label.
To break this bottleneck, researchers are moving away from pure supervised learning. Instead, they are developing techniques that extract supervision from unlabeled data, synthesize realistic training examples, and leverage "weak" signals like points or noisy automated labels.
Consistency as a Supervisory Signal
If we don't have a ground-truth label for an image, how can we tell a model it is wrong? The core intuition of semi-supervised learning (SSL) in this field is consistency. If you perturb an image—by adding noise, rotating it, or viewing it through a different "task" lens—the model's prediction should remain stable.
Jieneng Chen and colleagues demonstrated this through Dual-task Consistency. Instead of just asking a network to classify pixels, they forced it to simultaneously perform geometric level-set regression. By requiring these two mathematically distinct tasks to produce the same boundary, the model learns the underlying anatomy of the heart or brain without needing a label for every training image.
Similarly, Chunhui Zhao and Zihang Liu introduced Geometry-aware Consistency, which uses distance maps to penalize geometric inconsistencies. This forces the AI to "understand" the physical continuity of organs, allowing models to reach high performance using only 10% of the usual labeled data.
Synthesizing the Clinical Reality
When real labels are scarce, researchers build their own. This goes beyond simple rotations; it involves generating entirely new, anatomically plausible images.
One approach is to use generative models to "fill in" the gaps. Xinrong Hu and Yiyu Shi developed AugPaint, which uses latent diffusion models to inpaint new backgrounds around existing labels. By keeping the labeled organ intact but varying the surrounding tissue and noise, they create a nearly infinite variety of training pairs.
Other researchers, like Amlan Kar and Nishant Ravikumar, use Federated Simulation (Fed-Sim). This framework disentangles human anatomy from scanner-specific noise, allowing multiple hospitals to collaboratively synthesize 3D CT scans and labels without ever sharing sensitive patient data. For rare cases like lung nodules, Siqi Liu showed that decomposing images into shape and texture allows for the synthesis of "hard" cases that rarely appear in standard datasets, significantly boosting detection accuracy.
Learning from Weak and Noisy Signals
Not all labels are created equal. Tracing a 3D tumor pixel-by-pixel is slow, but clicking the center of a lesion or drawing a line is fast.
Yuhan Xie and Changzhen Qiu explored a point-line weak supervision paradigm for meniscus segmentation. By combining these "cheap" labels with self-supervised pre-training (using Masked Autoencoders), they achieved results comparable to full supervision.
Another strategy is to "bootstrap" from existing, imperfect tools. Abhijit Guha Roy developed QuickNAT and SD-Net by first training on "noisy" labels generated by traditional, non-AI software. The model then undergoes Error Corrective Boosting on a tiny set of high-quality manual labels to refine its accuracy. This two-step process allows the model to learn the general structure from the noise and the fine details from the experts.
The Rise of Few-Shot and In-Context Learning
The most extreme version of this research direction is "one-shot" or "few-shot" learning, where a model generalizes from just a handful of examples.
Devavrat Tomar proposed a generative style transfer framework that can train a segmentation model using only one labeled atlas. By treating other unlabeled images as "styles" to be applied to the atlas, the system synthesizes a diverse training set from a single source.
Taking inspiration from Large Language Models, Neelesh Kumar and Oya Aran demonstrated Visual In-Context Learning for eczema segmentation. By simply showing a generalist model two example images at inference time, it outperformed specialized U-Nets trained on hundreds of images. This suggests a future where medical AI isn't "trained" for every new hospital, but rather "prompted" with a few expert examples.