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

Zixuan Pan1, Jun Xia1, Zheyu Yan1, Guoyue Xu1, Yawen Wu1, Zhenge Jia2, Jianxu Chen3, Yiyu Shi1
1University of Notre Dame, 2Shandong University 3Leibniz-Institut für Analytische Wissenschaften - ISAS e.V.
arXiv Code Data

Abstract

Anomaly detection in brain MRI is crucial for identifying abnormalities such as tumors. Reconstruction-based methods, particularly those utilizing autoencoders, have been prevalent in this area. While much of the existing work has focused on meticulously designing models and algorithms, we revisit the problem of reconstruction-based anomaly detection in brain MRI from an image quality assessment perspective—an underexplored aspect in this field. We propose a fusion quality loss function that combines Structural Similarity Index Measure loss with ℓ1 loss, offering a more comprehensive evaluation of reconstruction quality. Additionally, we introduce a dataset pre-processing strategy that enhances the average intensity ratio between normal and abnormal regions, improving the distinction of anomalies. Our method achieves state-of-the-art performance on the BraTS21 (T2, FLAIR) and MSULB datasets, demonstrating significant improvements (>10%) in both Dice coefficient (DICE) and Area Under the Precision-Recall Curve (AUPRC) compared to existing methods. This work highlights the importance of comprehensive image quality assessment in medical anomaly detection and provides a new perspective for future research in this field.

model image.

Overview of our reconstruction-based anomaly detection method using a denoising autoencoder (DAE) as the reconstruction model. During training, the normal dataset Xn is augmented with the proposed AIR enhancement pre-processinig strategy and corrupted to form the noisy normal dataset Xn′ using simplex noise. The model is then trained by denoising Xn and minimizing the fusion quality loss LFQ between the reconstruction ^ Xn and the original normal dataset Xn. During inference, the abnormal test dataset Xa undergoes the same process. The anomalies in Xa are expected to be poorly reconstructed, resulting in higher values in the LFQ-based anomaly map. The final anomaly map is thresholded for segmentation, with performance measured by DICE and AUPRC.

Results

Table Conversion
Comparison with baselines in terms of DICE and AUPRC on BraTS21 and MSLUB using T2 modality in a cross-dataset setting. The model is trained on the IXI dataset, which contains only healthy samples. The best results for a given metric/dataset are bolded, and the second-best are underlined.
Method BraTS21 (T2) MSLUB (T2)
DICE [%] AUPRC [%] DICE [%] AUPRC [%]
Thresh [1] 19.69 20.27 6.21 4.23
AE [2] 32.87±1.25 31.07±1.75 7.10±0.68 5.58±0.26
VAE [2] 31.11±1.50 28.80±1.92 6.89±0.09 5.00±0.40
SVAE [3] 33.32±0.14 33.14±0.20 5.76±0.44 5.04±0.13
DAE [4] 37.05±1.42 44.99±1.72 3.56±0.91 5.35±0.45
f-AnoGAN [5] 24.16±2.94 22.05±3.05 4.18±1.18 4.01±0.90
DDPM [6] 40.67±1.21 49.78±1.02 6.42±1.60 7.44±0.52
mDDPM [7] 51.31±0.66 57.09±0.94 8.08±0.70 9.06±0.62
pDDPM [8] 49.41±0.66 54.76±0.83 10.65±1.05 10.37±0.51
pDDPM-IQA (ours) 59.45±0.37 62.99±0.37 12.93±0.67 11.51±0.50



results image

Qualitative visualization on the BraTS21 test set. Columns 2-5 show anomaly maps from different methods for three samples.

BibTeX

@misc{pan2024rethinkingmedicalanomalydetection,
      title={Rethinking Medical Anomaly Detection in Brain MRI: An Image Quality Assessment Perspective},
      author={Zixuan Pan and Jun Xia and Zheyu Yan and Guoyue Xu and Yawen Wu and Zhenge Jia and Jianxu Chen and Yiyu Shi},
      year={2024},
      eprint={2408.08228},
      archivePrefix={arXiv},
      primaryClass={eess.IV},
      url={https://arxiv.org/abs/2408.08228},
}