Multi-Instance Deep Learning: Discover Discriminative Local Anatomies for Bodypart Recognition

Area of Science:

• Medical Image Analysis
• Computer Vision
• Machine Learning

Background:

• Discriminative information for image recognition, including medical imaging, often resides in local image patches.
• Identifying specific body parts from transversal slices relies on localized anatomical features, such as the mediastinum region.

Purpose of the Study:

• To develop a multi-stage deep learning framework for image classification, specifically for bodypart recognition in medical images.
• To automatically discover discriminative and non-informative local image regions for improved classification accuracy.
• To learn an image-level classifier leveraging these identified local regions.

Main Methods:

• A two-stage learning scheme involving a convolutional neural network (CNN) trained in a multi-instance learning fashion during a pre-train stage.
• Extraction of discriminative and non-informative local patches from training slices using the pre-trained CNN.
• A boosting stage where the pre-learned CNN is enhanced using the identified local patches for final image classification.

Main Results:

• The proposed framework successfully identifies discriminative local patches automatically, eliminating the need for manual annotation.
• The CNN model, by focusing on discriminative local appearances, achieved higher accuracy compared to models relying on global image context.
• Validation on synthetic and large-scale CT datasets demonstrated superior performance over state-of-the-art methods, including standard deep CNNs.

Conclusions:

• The developed multi-stage deep learning framework effectively performs bodypart recognition by automatically discovering and utilizing discriminative local image features.
• This approach offers a significant advancement in medical image analysis by enhancing classification accuracy through intelligent patch selection.
• The method's ability to learn without manual annotation makes it a valuable tool for large-scale medical image datasets.