DeepScope: Nonintrusive Whole Slide Saliency Annotation and Prediction from Pathologists at the Microscope

Area of Science:

• Digital Pathology
• Medical Machine Learning
• Computational Pathology

Background:

• Digital pathology generates massive whole-slide image data, ideal for machine learning.
• Lack of image-level annotations hinders supervised learning applications.
• Current FDA regulations require primary diagnosis from glass slides, not digital images.

Purpose of the Study:

• To develop an end-to-end framework for nonintrusive annotation of digital slides.
• To overcome the limitations of manual labeling in digital pathology.
• To enable large-scale medical machine learning using routine clinical data.

Main Methods:

• Utilizing 3D-printed camera mounts to video record the glass-slide diagnosis process.
• Registering video frames to digital slides and estimating motion/observation time.
• Generating spatial and temporal saliency maps for annotation.

Main Results:

• A convolutional neural network trained on saliency maps achieved 85.15% accuracy in bladder and 91.40% in prostate cancer detection.
• Cross-tissue prediction accuracy reached 75.00% (prostate to bladder).
• Area Under the Receiver Operating Characteristic curve (AUROC) was 0.79±0.11 for bladder and 0.96±0.04 for prostate when training on one patient and testing on another.

Conclusions:

• The developed framework effectively generates annotations from routine pathologist workflows.
• This approach facilitates large-scale supervised learning in digital pathology.
• The tool demonstrates high accuracy in detecting diagnosis-relevant salient regions across different tissues and patients.