Object-based feedback attention in convolutional neural networks improves tumour detection in digital pathology

Area of Science:

• Computer Vision
• Computational Neuroscience
• Medical Imaging Analysis

Background:

• Human visual attention prioritizes behaviorally relevant information by allocating limited processing resources.
• Existing image recognition systems often lack mechanisms to dynamically focus on informative image regions.

Purpose of the Study:

• To develop an image recognition system inspired by biological vision that guides attention to salient locations within large images.
• To improve classification accuracy and provide explainability in image analysis through a feedback-driven attention mechanism.

Main Methods:

• A novel attention model employing saccade-like motions to sequentially process informative image regions.
• Utilizing feedback activations to highlight salient features and enable non-linear interactions between feedforward and backward passes.
• Applying the model to histopathology patch images from colorectal cancer whole slide images (WSIs).

Main Results:

• The attention system demonstrated improved classification ability compared to standard convolutional neural networks (CNNs).
• A 3.5% accuracy improvement (p < 0.001) was achieved on 59,057 9-class histopathology patches.
• The saccade implementation reached 93.23% agreement for tumor tissue classification, exceeding inter-pathologist agreement.

Conclusions:

• The proposed feedback-driven attention model enhances image recognition accuracy and explainability.
• This biologically inspired approach offers a significant advancement for analyzing large-scale scientific images, particularly in digital pathology.
• The method is adaptable for applications in diverse scientific fields requiring large-scale image analysis.