An interpretable classifier for high-resolution breast cancer screening images utilizing weakly supervised...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Natural image neural networks are often unsuitable for high-resolution medical images.
• Medical image analysis requires specialized architectures to handle unique properties like small regions of interest.

Purpose of the Study:

• To propose a novel neural network model tailored for medical image analysis.
• To address the limitations of existing models in processing high-resolution medical data.
• To improve the accuracy and efficiency of detecting lesions in medical images.

Main Methods:

• A two-stage network identifies informative regions using a low-capacity network, then analyzes details with a higher-capacity network.
• A fusion module integrates global and local information for prediction.
• The model is trained with image-level labels and generates pixel-level saliency maps, eliminating the need for lesion segmentation during training.

Main Results:

• Achieved an AUC of 0.93 on the NYU Breast Cancer Screening Dataset, outperforming ResNet-34 and Faster R-CNN.
• Attained an AUC of 0.858 on the CBIS-DDSM dataset, comparable to state-of-the-art methods.
• Demonstrated 4.1x faster inference and 78.4% less GPU memory usage compared to ResNet-34.
• Outperformed radiologists in a reader study with a 0.11 AUC margin.

Conclusions:

• The proposed neural network model effectively analyzes medical images, particularly in mammography interpretation.
• The model offers significant improvements in efficiency and performance over existing methods.
• This approach has the potential to enhance diagnostic capabilities in medical image analysis.