Diffusion model boosts tongue image augmentation for colorectal cancer detection

In a striking fusion of traditional Chinese medicine diagnostics and cutting-edge artificial intelligence, researchers in China have unveiled a generative AI system capable of producing synthetic tongue images that could help sharpen computer-aided detection of colorectal cancer, one of the world’s most common malignancies. The new framework, known as DTMG-Net, addresses one of the most stubborn bottlenecks in medical AI: the chronic shortage of large, well-labeled clinical image datasets.

The study, published as an open-access research article in BMC Medical Imaging, was led by Lanlan Li, Yu Zeng and Ziyue Wang—joint first authors—along with colleagues at Fuzhou University, Sun Yat-sen University’s Sixth Affiliated Hospital in Guangzhou, and Guangdong Second Provincial General Hospital. Their central insight is deceptively simple: if real clinical images are scarce, why not synthesize realistic, diverse, diagnostically relevant ones? The difficulty, as any practitioner of medical image generation knows, lies in making synthetic images that are simultaneously faithful to the underlying pathology and varied enough to genuinely help a classifier learn.

Colorectal cancer, often abbreviated CRC, is routinely screened through tools such as the fecal immunochemical test and endoscopic examination. In parallel, tongue diagnosis—a pillar of traditional East Asian medicine—has attracted growing scientific interest because changes in tongue color, coating, texture and shape can correlate with systemic disease states. Deep-learning classifiers trained on tongue photographs have shown promise in distinguishing CRC patients from healthy controls. But these models are voracious consumers of data, and clinical tongue image collections are typically small, imbalanced and expensive to curate. Conventional augmentation techniques—rotations, flips, brightness shifts, crops—merely rearrange existing pixels without creating new pathological variation, and they cannot mimic the complex, clinically meaningful differences that separate diseased from healthy tongues.

The research team’s answer was to build an improved diffusion-based generative model. Diffusion models, which learn to create images by reversing a gradual noising process, have taken the computer vision world by storm in recent years. The researchers started from the Denoising Diffusion Implicit Model, or DDIM, a fast-sampling variant of the widely used DDPM family, and then substantially re-engineered its denoising backbone, a U-Net neural network, in two important ways.

The first innovation is the embedding of a DeepSeek Mixture-of-Experts, or DeepSeekMoE, module into the denoising U-Net. Mixture-of-experts architectures are a class of sparse neural networks in which a gating mechanism routes each input to specialized subnetworks—”experts”—rather than pushing all data through a single monolithic set of weights. In DTMG-Net, this routing operates adaptively across the diffusion timesteps: early in the reverse process, when an image is nearly pure noise, the model faces a very different task than in later steps, when fine anatomical and textural details must be resolved. By allowing different expert subnetworks to specialize at different stages of denoising and at different scales of pathological features, the model gains a more nuanced capacity to represent the multi-scale structure of tongue imagery, from the global shape of the tongue body down to the subtle lesions that matter diagnostically.

The second innovation is a multi-scale dilated attention block, or MSDA, placed at the bottleneck of the U-Net—the point in the network where spatial resolution is lowest and semantic abstraction is highest. Dilated attention applies attention mechanisms across receptive fields that have been expanded with dilation, letting the network capture both long-range, global structure (the overall geometry and color distribution of the tongue) and fine-grained local texture (coating patterns, cracks, and lesion-like features) without the prohibitive memory cost of full-resolution attention. The joint capture of global form and local detail is precisely what makes a synthetic tongue image look plausible both at a glance and under the scrutiny of a trained classifier.

A third element of the design targets a well-known failure mode of generative models: mode collapse, in which a generator learns to produce a narrow set of safe, repetitive outputs rather than exploring the full diversity of the data distribution. The team designed a joint loss function that combines the standard noise-estimation objective of diffusion training with an intra-sample diversity regularization term. In effect, the model is rewarded not only for reconstructing realistic images but also for producing variations that differ meaningfully from one another, expanding the visual richness of the synthetic tongue samples without sacrificing fidelity.

Quantitatively, the authors evaluated their generated images using two standard metrics in generative modeling. The Fréchet Inception Distance, or FID, measures the statistical distance between the distribution of generated images and that of real images—lower is better. The Inception Score, or IS, rewards generators for producing images that are both classifiable and diverse—higher is better. Benchmarked against a variational autoencoder (VAE), a deep convolutional generative adversarial network (DCGAN), the PNDM diffusion model, and the original DDIM, DTMG-Net achieved FID values of 73.83 for CRC tongue samples and 58.99 for healthy control samples. The researchers are candid that these absolute FID values remain relatively high—a reflection of the exceptionally small-sample setting of tongue image data, which makes distribution matching inherently difficult. What matters, they argue, is that DTMG-Net attained the smallest distribution discrepancy of any generative model compared in the study under these challenging conditions.

Ablation experiments—experiments in which individual components are removed to test their contribution—confirmed that both the DeepSeekMoE module and the MSDA block independently improved generation performance, and that the diversity constraint raised the Inception Score without a meaningful degradation in FID. That balance is the whole point: a generator that is realistic but repetitive teaches a classifier little; one that is diverse but implausible can actively mislead it.

The decisive test, however, was downstream. The team augmented training sets with DTMG-Net synthetic images and trained four mainstream classification backbones: WideResNet, ResNet50, MedMamba and the Vision Transformer, or ViT. These architectures span the modern deep-learning landscape, from convolutional workhorses to state-space hybrids and transformer-based models. Across these backbones, training sets enriched with DTMG-Net generated images generally produced higher numerical values of the area under the ROC curve (AUC), F1-score, and accuracy than training sets augmented with images from the competing generative strategies. The improvement was consistent in direction though, as the authors note, “numerical” in character—measured within the constraints of their available dataset—yet the pattern held across diverse architectures, which strengthens the case that the synthetic data carries genuine diagnostic signal rather than dataset-specific noise.

The clinical implications are noteworthy. Screening for colorectal cancer remains imperfect: adherence to colonoscopy is limited, and non-invasive tests such as the fecal immunochemical test have well-documented sensitivity constraints. A supplementary, entirely non-invasive modality based on an ordinary photograph of the tongue—an examination that costs nothing, causes no discomfort, and requires no laboratory infrastructure—could, if validated at scale, complement existing screening pathways, particularly in resource-limited settings or in large-scale community screening campaigns where endoscopic capacity is scarce.

The study also exemplifies a broader and rapidly growing trend in medical AI: the use of generative augmentation to rescue small clinical datasets. Rare diseases, niche imaging modalities and traditional medicine modalities alike suffer from data scarcity that prevents deep networks from reaching their potential. Generative approaches of the kind embodied in DTMG-Net offer a path forward that conventional augmentation cannot: the creation of new, plausible, pathological variation rather than mere geometric transformations of existing images. The researchers emphasize that their scheme requires no complex preprocessing, which lowers the barrier to deployment in practical clinical workflows.

The work was supported by the General Program of the National Natural Science Foundation of China, and was conducted under ethical approval from the Medical Ethics Committee of the Sixth Affiliated Hospital of Sun Yat-sen University, with informed consent from all subjects and anonymized research data. The article was published open access on 8 September 2026, having been received on 30 June and accepted on 31 August of that year, and is available under a Creative Commons license.

Caveats remain, as they always do at this stage of translational research. FID values in the tens indicate that synthetic tongue images are still not statistically indistinguishable from real ones, and downstream performance gains, while consistent, were evaluated on small-scale datasets. Prospective, multi-center validation with large, diverse patient cohorts will be needed before any tongue-image-based CRC screening tool reaches the clinic. Privacy and data-governance considerations around synthetic medical imagery will also demand careful attention.

Nevertheless, DTMG-Net offers a compelling demonstration that ideas flowing from the frontiers of generative AI—sparse mixture-of-experts routing, dilated multi-scale attention, diversity-aware training objectives—can be purpose-built for problems as specific and as human as reading a tongue. In doing so, the study points toward a future in which the ancient diagnostic art of tongue inspection is augmented, rather than replaced, by machines trained to see patterns that even experienced clinicians might miss. It also adds to the accumulating evidence that generative augmentation is becoming an indispensable tool in the medical AI toolkit, turning small, hard-won clinical datasets into the training corpora that modern deep networks demand.