Abstract
Unsupervised anomaly detection in medical imaging offers a scalable alternative to supervised methods that require extensive expert labeling. This paper presents a hybrid approach for mammography screening that combines the feature extraction capabilities of Autoencoders (AEs) with the rigorous probabilistic inference of Sum-Product Networks (SPNs). By modeling the distribution of healthy tissue in a compressed latent space, the proposed system provides precise likelihood estimations for identifying anomalies. Evaluations on the DDSM and CBIS-DDSM datasets demonstrate that the $\beta$VAE-RATSPN hybrid achieves a pixel-wise AUC of 0.88 for mass detection, significantly outperforming standalone generative models and the state-of-the-art f-AnoGAN.
Introduction
Automated cancer screening is essential for managing the increasing workload of radiologists and ensuring early detection of breast cancer. Traditional supervised learning models require thousands of pixel-level annotations, which are costly and time-consuming to obtain. Unsupervised anomaly detection addresses this by training models exclusively on "healthy" data to learn the distribution of normal tissue. Any significant deviation from this learned norm is flagged as a potential anomaly.
While standard tools like Generative Adversarial Networks (GANs) and Autoencoders (AEs) are effective at feature extraction, they often struggle to provide mathematically exact likelihood scores, typically relying on noisy reconstruction errors. This paper proposes integrating SPNs to provide a more reliable probabilistic framework for anomaly detection.
Methodology
The proposed pipeline consists of two primary stages: dimensionality reduction and density estimation.
1. Dimensionality Reduction
The system uses an Autoencoder to compress 64x64 pixel image patches into a low-dimensional latent representation ($Z_n$). Three variants were evaluated: Convolutional AEs (CAE), $\beta$-Variational AEs ($\beta$VAE), and Vector Quantized VAEs (VQVAE). The Autoencoder serves as a noise filter, allowing the subsequent stage to focus on core structural features.
2. Density Estimation
The latent representations are fed into a Random and Tensorized Sum-Product Network (RAT-SPN). Unlike standard neural networks, SPNs are probabilistic graphical models that allow for exact and efficient inference. They calculate the actual probability (likelihood) that a specific latent code belongs to the "healthy" distribution.
Experimental Setup
The models were trained on the DDSM (Digital Database for Screening Mammography) using 695 healthy exams, from which 120 patches per image were extracted. Evaluation was performed on the CBIS-DDSM (Curated Breast Imaging Subset), which includes ground truth labels for 79 scans with masses and 30 scans with calcifications.
Results and Discussion
The hybrid models were compared against standalone AEs and the f-AnoGAN benchmark. A key finding was that the SPN significantly increases the "gap" between the scores of healthy and anomalous tissue, facilitating more effective thresholding.
- Mass Detection: The $\beta$VAE-RATSPN was the top performer, achieving a pixel-wise Area Under the Curve (AUC) of 0.88, compared to 0.80 for the standalone $\beta$VAE and 0.86 for f-AnoGAN.
- Localization: The hybrid models demonstrated improved segmentation accuracy for masses, as measured by the Hausdorff distance.
- Latent Space Modeling: The RAT-SPN performed better with continuous latent spaces (CAE and $\beta$VAE) than with the discrete space of the VQVAE.
Conclusion and Future Work
The integration of SPNs with generative models provides a robust framework for unsupervised anomaly detection in mammography. While the system excels at mass detection, performance on small calcifications remains a challenge due to their similarity to dense breast tissue in low-resolution patches. Future research will focus on increasing input resolution and developing methods to account for varying breast tissue densities to improve the detection of subtle anomalies.
Summary updated