Source-linked AI summary
Unsupervised Lesion Detection via Image Restoration with a Normative Prior
Xiaoran Chen, Suhang You, Kerem Can Tezcan, Ender Konukoglu
TL;DR
Unsupervised lesion detection remains challenging and currently trails supervised approaches, despite its value for robust algorithm development and clinical practice. This paper frames detection as MAP-based image restoration using a learned normative prior, achieving better AUC and Dice scores than existing methods while remaining limited for small or unobvious lesions.
Problem
Unsupervised lesion detection remains challenging and currently yields lower accuracy than supervised approaches, despite its value for robust algorithm development and clinical practice.
Method
The method estimates a normative prior with deep-learning latent-variable models, then restores lesion-containing images using MAP estimation and detects lesions from restoration differences.
Results
The proposed method achieved better AUC and Dice scores, significantly outperforming existing unsupervised methods on brain lesion datasets.
Takeaways & Limitations
MAP-based restoration improves the state of unsupervised lesion detection and moves it toward viability as an alternative to supervised approaches.
Takeaways & Limitations
The method has limited capability for detecting small and unobvious lesions, while large deformations may still be detected as abnormalities.
Abstract
from arXiv · showhide
Unsupervised lesion detection is a challenging problem that requires accurately estimating normative distributions of healthy anatomy and detecting lesions as outliers without training examples. Recently, this problem has received increased attention from the research community following the advances in unsupervised learning with deep learning. Such advances allow the estimation of high-dimensional distributions, such as normative distributions, with higher accuracy than previous methods.The main approach of the recently proposed methods is to learn a latent-variable model parameterized with networks to approximate the normative distribution using example images showing healthy anatomy, perform prior-projection, i.e. reconstruct the image with lesions using the latent-variable model, and determine lesions based on the differences between the reconstructed and original images. While being promising, the prior-projection step often leads to a large number of false positives. In this work, we approach unsupervised lesion detection as an image restoration problem and propose a probabilistic model that uses a network-based prior as the normative distribution and detect lesions pixel-wise using MAP estimation. The probabilistic model punishes large deviations between restored and original images, reducing false positives in pixel-wise detections. Experiments with gliomas and stroke lesions in brain MRI using publicly available datasets show that the proposed approach outperforms the state-of-the-art unsupervised methods by a substantial margin, +0.13 (AUC), for both glioma and stroke detection. Extensive model analysis confirms the effectiveness of MAP-based image restoration.
1. Introduction
Unsupervised lesion detection learns healthy anatomy as a normative reference and identifies deviations without lesion-specific training examples. The proposed method treats detection as MAP-based image restoration to address false positives from prior projection.
- Manual lesion detection identifies regions that deviate from normal healthy anatomy without searching for a specific lesion type.
- Supervised algorithms achieve strong performance on specified lesion types but have limited applicability to unseen lesions.
- Unsupervised methods learn healthy-anatomy priors and detect lesions as areas that disagree with the learned normative distribution.
- Deep learning enables high-dimensional normative-distribution modeling through neural samplers and nonlinear latent-variable models such as VAEs.
- Prior projection can produce false positives because lesion intensities may substantially alter latent representations and cause healthy regions to deviate during reconstruction.
- The proposed approach formulates unsupervised detection as MAP image restoration using VAE or GMVAE priors and evaluates it on brain-tumor and stroke-lesion MRI datasets.
2. Method
The method learns a normative distribution of healthy anatomy with latent-variable models, then restores lesion-containing images using MAP estimation so normal anatomy is preserved while outliers are detected from restoration differences.
- Learning the normative distribution: Latent-variable models represent high-dimensional images through a lower-dimensional latent space, with P(X|z) mapping latent variables back to image space.The latent dimension M is typically much smaller than the image dimension N.
- Learning the normative distribution: VAE approximates the posterior with a neural network and optimizes the evidence lower bound, while GMVAE uses a Gaussian-mixture latent prior for more expressive distributions.Both models are trained on healthy MR images to estimate the normative distribution.
- MAP-based restoration: The restoration model treats a lesion-containing image as a healthy image plus an additive lesion component and aims to retain normal anatomy while changing only the outlier region.For lesion-free images, the model assumes the observed and clean images are identical.
- MAP-based restoration: MAP estimation balances data consistency with the normative prior, penalizing deviations from the observed image while favoring restored images that fit healthy anatomy.The optimum assigns poorly fitting lesion content to the outlier component while keeping normal anatomy fixed.
- Pixel-wise detection: After optimization, the restored image is compared with the observation to produce a pixel-wise lesion map, with absolute differences capturing both hypo-intense and hyper-intense lesions.Binary segmentation thresholds the continuous difference map using a false-positive-rate limit estimated from healthy validation images.
- Pixel-wise detection: Healthy images can still change during restoration because the learned log-probability, ELBO, or anatomical preferences may imperfectly represent the normative distribution.This motivates tuning restoration parameters using healthy validation images.
3.1. Datasets & Preprocessing
The study uses healthy CamCAN scans to train and tune normative models, and BRATS17 and ATLAS scans to evaluate lesion detection after standardized preprocessing and slice extraction.
- Datasets: BRATS17 provides T1- and T2-weighted scans from 285 subjects with brain tumors, including high- and low-grade gliomas, and evaluation uses T2-weighted images.The dataset version was published in 2017.
- Datasets: Prior models trained on CamCAN T2-weighted images support BRATS17 evaluation, while T1-weighted prior models support ATLAS evaluation.This assigns modality-matched normative models to the two lesion datasets.
- Preprocessing: Preprocessing includes skull stripping, MNI registration, histogram matching across CamCAN subjects, and pixel-intensity normalization.These steps are applied before training the normative models.
- Preprocessing: All computations use independent 2D transversal slices of size 200 × 200, while detection runs slice-wise and metrics are computed subject-wise.Slices without brain structures and excessive background are excluded.
3.2. Implementation Details
The implementation evaluates MAP-based restoration alongside prior-projection baselines using VAE and GMVAE normative models, with ROC/AUC and FPR-controlled Dice metrics. The models use convolutional encoder–decoder architectures and fixed restoration optimization settings.
- Normative models: Two normative prior sets were trained with VAE and GMVAE models, separately for T1-weighted and T2-weighted images using CamCAN data.BRATS17 and ATLAS images were excluded from prior-model training.
- Network architecture: The VAE uses six down-sampling residual blocks, a symmetric decoder, and a 2×2×512 latent variable.Down-sampling uses stride 2, while up-sampling uses bilinear interpolation.
- Baselines: A U-Net with the same encoder–decoder structure and skip connections provides a supervised baseline trained with cross-entropy loss.The architecture was selected for experimental consistency rather than state-of-the-art leaderboard performance.
- Optimization: The proposed detector performs 500 restoration steps, using step size 5 × 10−3 initially and 3 × 10−3 thereafter.The first step size is used for 100 iterations; the strategy was selected by final MAP values rather than evaluation metrics.
- Baselines: The comparison includes VAE-256, VAE-128, AAE-128, and AnoGAN prior-projection methods with hyperparameters tuned on the study datasets.These methods are compared against the proposed VAE and GMVAE MAP-based restoration approaches.
- Evaluation: Performance is evaluated with AUC and Dice scores at 1%, 5%, and 10% false-positive-rate thresholds, plus a DSC AUC threshold baseline.AUC is extracted from ROC curves, while DSC values are computed subject-wise from thresholded detection maps.
3.3. Results
The MAP-based method outperformed unsupervised baselines on BRATS17 and ATLAS, while performance remained more difficult on ATLAS and for smaller lesions. Visual and threshold analyses further characterized detection quality and operating-point selection.
- Model performance: AUC exceeded 0.80 for every proposed prior model, whereas BRATS17 baselines achieved 0.70 or lower.The proposed MAP-based variants also yielded higher AUC on ATLAS than baseline methods.
- Model performance: At a 1% FPR limit on BRATS17, the proposed method reached mean DSC 0.34 with VAE and more than 0.20 with every GMVAE variation.Baseline methods remained below 0.10 in this conservative setting.
- Dataset differences: ATLAS DSC values were substantially lower than BRATS17 values because lesions were smaller and often had intensities similar to normal structures.The proposed methods still achieved substantially higher DSC than baselines for both prior terms.
- Prior-model comparison: GMVAE with c = 9 outperformed VAE on BRATS17 at 5% and 10% FPR, but this increase was not observed on ATLAS.The results did not identify a clear overall winner between VAE and GMVAE priors.
- Threshold analysis: Automatic thresholds produced DSC values similar to thresholds selected from ROC curves across methods, supporting the automatic thresholding procedure.ROC-based threshold selection used lesion annotations and was therefore not unsupervised, whereas the automatic procedure was designed for unsupervised use.
- Visual analysis: Restoration substantially reduced abnormal intensities, enabling difference maps and thresholded detections that matched large BRATS17 tumors more closely than small tumors.Small tumors were harder to detect and could be confused with abnormal-looking healthy tissue or missed.
- Visualizations: The ROC plots compare all methods on complete BRATS17 and ATLAS datasets, with the dashed line marking random detection.The visual examples show inputs, restored images, continuous maps, binary maps at 1%, 5%, and 10% FPR, and ground-truth segmentations.
3.4. Accuracy Analysis with Lesion Size
Detection accuracy generally improves with lesion size, especially on ATLAS, but lesion characteristics beyond size also influence performance. Small lesions are particularly difficult to detect reliably.
- Higher Dice scores are often obtained for subjects with larger lesions, with the tendency stronger on ATLAS than BRATS17.
- FPR is mostly 0.10–0.20 on BRATS17 and mostly below 0.10 on ATLAS.
- On ATLAS, lesions smaller than 25000 pixels mostly produce Dice scores below the best average Dice of 0.12.
- On BRATS17, lesions larger than approximately 25000 pixels can achieve Dice scores above the optimal average of 0.45, but scores range from 0.1 to 0.8.
- Lesion intensity, location, and shape may also affect detection accuracy in a complicated way.
3.5. 3D consistency in detection
The proposed model produces detection results that remain consistent across slices in 3D brain volumes.
- Detected lesions are consistent across slices in the evaluated 3D scans.
3.6. Model Analysis with GMVAE prior
Model analysis examined hyperparameter sensitivity and restoration convergence for the GMVAE prior. Performance was generally robust to parameter choices, and iterative MAP restoration converged stably.
- The analysis varied data consistency weight λ, Gaussian mixture count c, and latent dimension dimz for the GMVAE model.
- λ values selected using ϵ(λ) were close to those producing maximum AUC across prior terms and datasets.
- Selecting λ with ϵ(λ) yielded AUC values at most 0.01 below the maximum.
- GMVAE hyperparameters had minimal effect on AUC, while 512 latent dimensions tended to provide the highest DSC scores across FPR limits and cluster counts.
- MAP-restoration AUC increased sharply early and stabilized around 0.8 after roughly 500 iterations on BRATS17 and 300 iterations on ATLAS.
4. Conclusions
The paper proposes unsupervised lesion detection through iterative MAP restoration using a learned normative prior. Experiments and analysis support improved detection, robust parameter behavior, and stable convergence, while identifying limitations for small lesions and large deformations.
- The method restores images iteratively with MAP estimation using VAE- or GMVAE-based normative priors learned from healthy images.
- Figure 10 analyzes λ selection, AUC stability across λ, and AUC evolution during MAP optimization for the GMVAE prior.
- Detection on brain lesion datasets achieved better AUC and Dice scores than existing methods, with stable convergence and robust GMVAE parameter selection.
- The method has limited capability for detecting small and unobvious lesions.
- Large deformations in surrounding healthy structures may still be detected as abnormalities despite the data consistency term, although this was not observed experimentally.