The code for MOOD(MICCAI 2020)
Biomedical image computing is a valuable but challenging task, which still largely depends on the availability of annotated training sets. However, labeling specific training data in advance is not only time-consuming but also labor-intensive in real applications. Furthermore, the out-of-distribution samples make the biomedical image analysis task more difficult.
To tackle these problems, we formulate this challenge as an anomaly detection task that can be solved using a reconstruction strategy. By observing the discriminative reconstruction errors, the biomedical images with high reconstruction losses are most likely to be the abnormalities. Therefore, we adopt U-Net architecture, which has an encoder-decoder structure with skip connections, to reconstruct the biomedical image. Moreover, ensemble technology is performed to combine the texture feature extracted by Canny operator and the context information generated by mask strategy on the foreground.
The experimental results conducted on two standardized datasets given by MOOD Challenge indicate that our method is considerably superior.