DGR-MIL: Diverse Global Representation in Multiple Instance Learning

Histological whole slide image (WSI) processing techniques are critical for identifying and classifying tissues, particularly cancerous tissues. AI-aided methods for analyzing these WSIs have become increasingly important to enhance diagnostic precision and high throughput processing. However, WSIs have gigapixel resolution, leading to computational intractability with conventional deep learning methods.

One such AI-aided method used in histological WSI classification involves multiple instance learning (MIL), which is a powerful approach for weakly supervised learning. Unfortunately, current MIL methods focus on modeling correlation between instances and overlook diversity among instances or patches.

 

Researchers at Arizona State University in collaboration with researchers at Clemson University have developed a novel machine learning technique for histological whole slide image processing. This technique employs a MIL aggregation method based on diverse global representation (DGR-MIL), by modeling diversity among instances through a set of global vectors that serve as a summary of all instances. Under this MIL setting, the WSI is treated as a bag and all the patches within the image are treated as instances of the bag. This allows modeling of diversity among distances through a set of learnable global vectors to be more descriptive of the entire bag. This proposed technique outperforms the current MIL aggregation models by a substantial margin on the CAMELYON-16 and the TCGA-lung cancer datasets.

 

Potential Applications

Benefits and Advantages

For more information about this opportunity, please see

 

For more information about the inventor(s) and their research, please see