Anomaly Detection for Medical Images Using Heterogeneous Auto-Encoder.

Area of Science:

• Medical image analysis
• Artificial intelligence in healthcare
• Deep learning for anomaly detection

Background:

• Anomaly detection in medical imaging is crucial for reducing reliance on large labeled datasets.
• Existing pixel-wise self-reconstruction methods struggle with overfitting and undesirable reconstruction results.
• Challenges include learning identity mappings and handling pixel-wise differences in anomaly detection.

Purpose of the Study:

• To propose a novel heterogeneous Auto-Encoder (Hetero-AE) for improved medical anomaly detection.
• To address limitations of existing methods, such as overfitting and sensitivity to pixel-wise noise.
• To enhance the accuracy and interpretability of anomaly detection in diverse medical imaging modalities.

Main Methods:

• Utilized a heterogeneous Auto-Encoder (Hetero-AE) with a CNN encoder and a hybrid CNN-Transformer decoder.
• Introduced a multi-scale sparse Transformer block to balance feature dependency modeling and computational cost.
• Implemented multi-stage feature comparison to minimize noise from pixel-wise comparisons.

Main Results:

• Demonstrated effectiveness across four public datasets: retinal OCT, chest X-ray, brain MRI, and COVID-19.
• Achieved accurate detection of tumors in brain MRI and lesions in retinal OCT.
• Generated interpretable heatmaps for precise localization of abnormalities, aiding clinical diagnosis.

Conclusions:

• The proposed Hetero-AE model offers a robust solution for anomaly detection in various medical imaging modalities.
• The heterogeneous architecture and novel Transformer block effectively learn normal data characteristics and highlight anomalies.
• The method provides valuable tools for clinicians by accurately detecting and localizing abnormalities, supporting efficient diagnosis.