Hierarchical Deep Learning for Abnormality Classification in Mouse Skeleton Using Multiview X-Ray Images:...

Area of Science:

• Biomedical imaging
• Machine learning
• Computational biology

Background:

• Single-view anomaly detection lacks context for multi-label problems.
• Multi-view imaging offers richer contextual information for classification tasks.

Purpose of the Study:

• To evaluate the efficacy of multi-view (MV) image data with hierarchical learning for skeletal abnormality detection.
• To compare the performance of MV classification against single-view methods in a multi-label context.

Main Methods:

• Curated a specimen-wise MV dataset from 170,958 International Mouse Phenotyping Consortium (IMPC) images.
• Developed two hierarchical classification frameworks using ConvNeXT and convolutional autoencoder (CAE) backbones.
• Trained models at three hierarchical levels of increasing anatomical granularity.

Main Results:

• MV classification performed comparably to single views at the top hierarchy level (L1, mean AUC 0.95).
• Hierarchical models using MV data significantly improved classification at lower levels (L2 and L3) compared to single views (e.g., L2: DV 0.65, LV 0.76, MV 0.87; L3: DV 0.54, LV 0.59, MV 0.82).
• Both ConvNeXT and CAE architectures demonstrated the advantage of MV data for detecting specific skeletal abnormalities.

Conclusions:

• Multi-view image data combined with hierarchical learning is advantageous for skeletal abnormality detection.
• This approach effectively addresses challenges in multi-label anomaly detection by providing enhanced contextual information.