Source-linked AI summary
Diffusion Models for Medical Anomaly Detection
Julia Wolleb, Florentin Bieder, Robin Sandkühler, Philippe C. Cattin
TL;DR
The paper targets medical anomaly detection with limited pixel-wise annotations and proposes DDIM-based image translation guided toward the healthy class. Across BRATS2020 and CheXpert, it reports detailed, detail-consistent anomaly maps without changing the original training procedure or architecture.
Problem
Medical anomaly detection often lacks pixel-wise annotations, while existing GAN and autoencoder approaches can be difficult to train or preserve fine image details.
Method
The method combines deterministic DDIM noising and denoising with classifier guidance to translate diseased images into healthy images and compute their difference.
Results
The method successfully translates images from BRATS2020 and CheXpert into images without pathologies while changing anomalous regions and producing detailed anomaly maps.
Takeaways & Limitations
Detail-consistent translation is achieved without changing the original implementations’ loss function, architecture, or training procedure.
Abstract
from arXiv · showhide
In medical applications, weakly supervised anomaly detection methods are of great interest, as only image-level annotations are required for training. Current anomaly detection methods mainly rely on generative adversarial networks or autoencoder models. Those models are often complicated to train or have difficulties to preserve fine details in the image. We present a novel weakly supervised anomaly detection method based on denoising diffusion implicit models. We combine the deterministic iterative noising and denoising scheme with classifier guidance for image-to-image translation between diseased and healthy subjects. Our method generates very detailed anomaly maps without the need for a complex training procedure. We evaluate our method on the BRATS2020 dataset for brain tumor detection and the CheXpert dataset for detecting pleural effusions.
1 Introduction
The paper addresses medical anomaly detection when pixel-wise labels are difficult to obtain by using image-level supervision. It introduces DDIM-based translation from diseased images to healthy images for detailed anomaly mapping.
- Pixel-wise medical annotations are difficult to obtain, often unavailable, and can reflect annotator bias.
- Weakly supervised models use only image-level labels during training, avoiding the need for pixel-wise ground truth.
- The anomaly map is the difference between the diseased input and generated healthy output.
- The method trains a DDPM and binary classifier, then uses DDIM encoding and classifier-guided denoising to generate a healthy version of an unseen image.
- Prior autoencoder and GAN approaches can face reconstruction-detail problems, challenging training, or extensive hyperparameter tuning.
2 Method
The method combines diffusion-based noise encoding with deterministic DDIM denoising and classifier guidance to translate an input image toward the healthy class. The input and generated image are differenced to produce a pixel-wise anomaly map.
- Diffusion models progressively add noise across timesteps and train a U-Net with mean squared error to predict less-noisy images.During evaluation, iterative denoising generates an image from Gaussian noise.
- DDIM sets σ_t = 0, making the sampling process deterministic and allowing the generation process to be reversed for image encoding.
- The method encodes an input image to a noisy state x_L, then denoises it with classifier guidance toward the healthy class h.The classifier gradient s∇_{x_t} log C(h|x_t,t) updates the denoising prediction.
- The generated healthy image preserves the input’s basic structure, and its difference from the input defines the anomaly map.
- Algorithm 1 uses input image x, healthy label h, gradient scale s, and noise level L to output synthetic image x_0 and anomaly map a.
3 Experiments
The experiments evaluate the method on CheXpert lung X-rays and BRATS2020 brain MR slices. CheXpert includes healthy and pleural-effusion subjects, while BRATS2020 provides multimodal tumor data with pixel-wise labels.
- The DDPM and classifier are trained with 50,000 and 20,000 iterations, respectively, using PyTorch 1.7.1.
- The evaluation uses CheXpert and BRATS2020 medical-image datasets.
- CheXpert training uses 14,179 healthy subjects and 16,776 subjects with pleural effusions, with 200 images per class in the test set.
- BRATS2020 contains four-channel axial brain MR slices from T1, T2, FLAIR, and contrast-enhanced T1 sequences.
- CheXpert examples are evaluated with L = 500 and s = 100.
4 Results and Discussion
The method produces detailed healthy translations and anomaly maps while preserving input anatomy, but its quality depends on the noise level and gradient scale. Healthy inputs are reconstructed with differences close to zero.
- Compared with FP-GAN and VAE, the method preserves input details and generates realistic healthy images, producing detailed anomaly maps.The comparison covers CheXpert examples; competing methods either alter unrelated regions or fail to detect an anomaly.
- Deterministic DDIM noise encoding reconstructs input details more faithfully than DDPM sampling, which changes basic anatomy.The complete translation takes about 158s, and reducing the noise level or skipping timesteps degrades image quality.
- Hyperparameter Sensitivity: Average Dice and AUROC scores are evaluated across gradient scales s and noise levels L on the BRATS2020 test set.Dice uses average Otsu thresholding on anomaly maps, while AUROC is computed for diseased slices.
- Hyperparameter Sensitivity: Too-large noise levels destroy images, whereas too-small levels leave the model insufficient freedom to remove tumors.These effects are illustrated while fixing the gradient scale at s = 100.
- Translation of a Healthy Subject: For a healthy BRATS slice, the method reconstructs the image in detail and produces an anomaly map close to zero.The input and synthetic image differ only minimally in this evaluation.
5 Conclusion
The paper combines iterative DDIM noising and denoising with classifier guidance for weakly supervised medical anomaly detection. It reports detail-consistent translations on BRATS2020 and CheXpert without modifying the original loss, architecture, or training procedure.
- The proposed method combines iterative DDIM noising and denoising with classifier guidance for weakly supervised anomaly detection.
- The method translates diseased patient images into images without pathologies while changing only anomalous regions.
- Using the original implementations' loss function and training scheme makes training on other datasets straightforward.
- The method achieves excellent results on the BRATS2020 and CheXpert datasets.
1 Additional Results on the CheXpert Dataset
The supplied passage identifies additional results for diseased subjects on the CheXpert dataset using L = 500 and s = 100.
- Additional method results are shown for diseased subjects on the CheXpert dataset with L = 500 and s = 100.
2 Additional Results on the Brats2020 Dataset
The supplied passage identifies additional results for a healthy subject on the BRATS2020 dataset using L = 500 and s = 100.
- Additional method results are shown for a healthy subject on the BRATS2020 dataset with L = 500 and s = 100.