[Paper Review] Brain Tumor Anomaly Detection via Latent Regularized Adversarial Network

Nan Wang, Cheng‐Wei Chen+2 more|arXiv (Cornell University)|Jul 9, 2020

Anomaly Detection Techniques and Applications26 references9 citations

TL;DR

This paper proposes a semi-supervised brain tumor anomaly detection method using a latent regularized adversarial network that trains only on normal brain MRI scans. By jointly optimizing image and latent space reconstruction with singular value constraints, the model achieves superior anomaly detection performance, outperforming state-of-the-art supervised methods on BraTS, HCP, MNIST, and CIFAR-10 with AUC scores up to 0.994 on BraTS using latent space reconstruction error as the anomaly score.

ABSTRACT

With the development of medical imaging technology, medical images have become an important basis for doctors to diagnose patients. The brain structure in the collected data is complicated, thence, doctors are required to spend plentiful energy when diagnosing brain abnormalities. Aiming at the imbalance of brain tumor data and the rare amount of labeled data, we propose an innovative brain tumor abnormality detection algorithm. The semi-supervised anomaly detection model is proposed in which only healthy (normal) brain images are trained. Model capture the common pattern of the normal images in the training process and detect anomalies based on the reconstruction error of latent space. Furthermore, the method first uses singular value to constrain the latent space and jointly optimizes the image space through multiple loss functions, which make normal samples and abnormal samples more separable in the feature-level. This paper utilizes BraTS, HCP, MNIST, and CIFAR-10 datasets to comprehensively evaluate the effectiveness and practicability. Extensive experiments on intra- and cross-dataset tests prove that our semi-supervised method achieves outperforms or comparable results to state-of-the-art supervised techniques.

Motivation & Objective

• Address the challenge of data imbalance and scarcity of labeled brain tumor images in medical imaging.
• Overcome the limitations of fully supervised methods that require extensive labeled abnormal data and assume known lesion types.
• Develop a robust, generalizable anomaly detection framework capable of identifying unknown or rare brain lesions without prior exposure.
• Improve detection performance by focusing on latent space optimization rather than relying solely on pixel-level reconstruction error.
• Enable practical deployment in clinical settings where collecting diverse pathological brain images is costly and time-consuming.

Proposed method

• Proposes a semi-supervised GAN-based framework that trains exclusively on normal (inlier) brain MRI images.
• Introduces a latent space regularization via singular value decomposition (SVD) to constrain the latent representation and improve feature disentanglement.
• Jointly optimizes the generator and discriminator using multiple loss functions: reconstruction loss in image space and latent space, and adversarial loss.
• Uses the reconstruction error in the latent space (||z - z'||) as the anomaly score, which is more discriminative than pixel-level reconstruction error.
• Employs a multi-objective training objective to enhance the separability of normal and abnormal samples in the latent feature space.
• Applies the trained model at inference to detect anomalies by measuring the deviation of test samples in the latent space from the normal distribution.

Experimental results

Research questions

• RQ1 Can a semi-supervised anomaly detection model trained only on normal brain MRI scans effectively detect diverse and unknown brain tumor types?
• RQ2 Does constraining the latent space via singular value decomposition improve the discriminative power of the learned features for anomaly detection?
• RQ3 How does latent space reconstruction error compare to pixel-level reconstruction error in distinguishing normal from abnormal brain images?
• RQ4 To what extent does the proposed method generalize across different datasets, including BraTS, HCP, MNIST, and CIFAR-10?
• RQ5 Can the model maintain high performance when tested on cross-dataset scenarios with distribution shifts?

Key findings

• The proposed method achieves an AUC of 0.994 on the BraTS dataset using latent space reconstruction error, significantly outperforming pixel-level reconstruction (AUC 0.938).
• On MNIST, the latent-based anomaly score improves AUC from 0.868 to 0.937, a 7.3% relative gain, demonstrating enhanced feature-level discrimination.
• On CIFAR-10, the AUC increases from 0.617 (pixel-level) to 0.797 (latent-level), showing consistent improvement in feature representation quality.
• The model maintains strong performance across cross-dataset evaluations, with minimal degradation even under distribution shift, indicating robust generalization.
• Visualization of the latent space shows clear separation between normal (blue) and abnormal (red) samples when using SVD regularization, confirming improved feature disentanglement.
• The model outperforms or matches state-of-the-art supervised methods such as ceVAE, VAE, AE, VAE-H, and ELBO Gradient in slice-wise tumor detection on BraTS.