Gigapixel end-to-end training using streaming and attention.

Area of Science:

• Computational pathology
• Digital pathology
• Machine learning in medicine

Background:

• Hardware limitations prevent direct training of convolutional neural networks on gigapixel images.
• Existing weakly supervised learning methods often use multi-stage or patch-wise strategies, risking suboptimal feature extraction.

Purpose of the Study:

• To propose an end-to-end training method for gigapixel microscopic images using only slide-level labels.
• To overcome hardware limitations in training deep learning models on large-scale medical images.

Main Methods:

• Developed StreamingCLAM, a ResNet-34 encoder with an attention-gated classification head, using a streaming implementation of convolutional layers.
• Enabled end-to-end training on 4-gigapixel images with slide-level labels.

Main Results:

• Achieved a mean area under the ROC curve of 0.9757 for metastatic breast cancer detection (CAMELYON16).
• Successfully detected MYC-gene translocation in diffuse large B-cell lymphoma with a mean AUC of 0.8259.
• Demonstrated interpretability through the attention mechanism.

Conclusions:

• StreamingCLAM facilitates end-to-end training of gigapixel images, achieving performance comparable to fully supervised methods.
• The attention mechanism provides insights into model predictions, enhancing interpretability.
• This approach offers a viable solution for large-scale computational pathology tasks.