Rethinking Medical Anomaly Detection in Brain MRI: An Image Quality Assessment Perspective

TL;DR

The paper tackles brain MRI anomaly detection under limited abnormal labels by reframing reconstruction quality as an image quality assessment (IQA) problem. It introduces Fusion Quality Loss, which linearly combines $L_{SSIM}$ and $L_{\ell_1}$ with $\alpha=0.84$, and an AIR-based transformation to amplify divisive discrepancies between normal and abnormal regions. Empirically, applying this IQA approach to a strong DDPM-based anomaly detector ($pDDPM$) yields significant improvements in Dice and AUPRC on BraTS21 and MSLUB and generalizes to other DDPM variants. This work demonstrates that an IQA perspective and simple pre-processing can substantially boost medical anomaly detection performance, suggesting a new direction for metric design in this domain.

Abstract

Reconstruction-based methods, particularly those leveraging autoencoders, have been widely adopted for anomaly detection task in brain MRI. Unlike most existing works try to improve the task accuracy through architectural or algorithmic innovations, we tackle this task from image quality assessment (IQA) perspective, an under-explored direction in the field. Due to the limitations of conventional metrics such as l1 in capturing the nuanced differences in reconstructed images for medical anomaly detection, we propose fusion quality, a novel metric that wisely integrates the structure-level sensitivity of Structural Similarity Index Measure (SSIM) with the pixel-level precision of l1. The metric offers a more comprehensive assessment of reconstruction quality, considering intensity (subtractive property of l1 and divisive property of SSIM), contrast, and structural similarity. Furthermore, the proposed metric makes subtle regional variations more impactful in the final assessment. Thus, considering the inherent divisive properties of SSIM, we design an average intensity ratio (AIR)-based data transformation that amplifies the divisive discrepancies between normal and abnormal regions, thereby enhancing anomaly detection. By fusing the aforementioned two components, we devise the IQA approach. Experimental results on two distinct brain MRI datasets show that our IQA approach significantly enhances medical anomaly detection performance when integrated with state-of-the-art baselines.

Figures (4)

• Figure 1: Visualization of the anomaly maps generated by $\ell_{1}$ loss and SSIM loss from the same reconstruction. Calculating the reconstruction discrepancy with L1-metric cannot flag the large tumor area, while calculating with SSIM, from the same reconstruction, could identify the tumor area significantly better.
• Figure 2: Overview of our reconstruction-based anomaly detection method with the proposed fusion quality loss and AIR-based data transformation. During training, the normal dataset $X_n$ is augmented with the proposed AIR-based data transformation to enhance the divisive discrepancies, and corrupted to form the noisy normal dataset $X_{n}{'}$ using simplex noise. The model is then trained by denoising $X_{n}'$ and minimizing the fusion quality loss $L_{FQ}$ between the reconstruction $\hat{X_n}$ and the original normal dataset $X_n$. During inference, the abnormal test dataset $X_a$ undergoes the same process. The anomalies in $X_a$ are expected to be poorly reconstructed, resulting in higher values in the $L_{FQ}$-based anomaly map. The final anomaly map is thresholded for segmentation, with performance measured in terms of DICE and AUPRC.
• Figure 3: Qualitative visualization on the BraTS21 test set. Columns 2-5 show anomaly maps from different methods for three samples.
• Figure 4: Ablation Study Results.