Summary of Challenges of Developing Models for Gigapixel-Scale Pathology | TransformX 2022

00:00:00 - 00:20:00

This video discusses the challenges of developing models for gigapixel-scale pathology, including the need to carefully choose samples during training and the potential for model shortcuts. A new memory-based approach is introduced that may help overcome some of these challenges. However, the approach is not without its own problems, including speed and compute issues, and a tendency to overfit.

• 00:00:00 Nathan Silverman will be speaking about the challenges of developing models for gigapixel scale pathology, including the difficulties of loading pathology images onto GPUs and the sheer scale of each image. He will also discuss the three ways to deal with these images.
• 00:05:00 The authors describe two methods for training machine learning models on gigapixel scale images: first, by splitting the image into tiles and training the model on each tile; and second, by using tiling, in which the pathologist labels each tile as either containing a tumor or being cancer free. They also describe a way to combine multiple predictions from a machine learning model to produce a single patient level prediction. While this approach is successful in training models on gigapixel scale images, there are a number of limitations, including the cost and time required to label each tile, the lack of spatial context in some cases, and the difficulty in making predictions for individual patients.
• 00:10:00 The video discusses the challenges of developing models for gigapixel-scale pathology, including the need to carefully choose samples during training and inference, and the potential for model shortcuts. Attention-based multiple instance learning can help overcome some of these challenges.
• 00:15:00 This YouTube video discusses the challenges of developing models for gigapixel-scale pathology, and introduces a new memory-based approach that achieves better performance without re-running computations.
• 00:20:00 This video discusses the challenges of developing models for gigapixel-scale pathology, and introduces a memory-based approach that is promising. However, the approach has speed and compute issues, and is prone to overfitting.