Reverse active learning based atrous DenseNet for pathological image classification.

Area of Science:

• Computational pathology
• Medical image analysis
• Deep learning

Background:

• Deep learning models are increasingly applied to medical image analysis.
• Pathological image analysis faces challenges like gigapixel resolution and limited annotation capabilities.
• Existing deep learning networks struggle with multiscale feature extraction in pathological images.

Purpose of the Study:

• To propose a novel training strategy and network architecture for pathological image classification.
• To address challenges in high-resolution pathological image analysis and improve classification accuracy.
• To enhance the performance of deep learning models on datasets with mislabeled patches.

Main Methods:

• Developed a deep-reverse active learning (DRAL) strategy to identify and remove mislabeled patches from training data.
• Proposed an atrous DenseNet (ADN) integrating atrous convolutions with dense blocks for multiscale feature extraction.
• Applied DRAL and ADN to pathological image classification tasks.

Main Results:

• The DRAL + ADN framework achieved excellent performance on three pathological datasets (BACH, CCG, UCSB).
• Patch-level average classification accuracies reached 94.10% (BACH), 92.05% (CCG), and 97.63% (UCSB).
• DRAL effectively improved classification accuracy of established networks like VGG-16 and ResNet.

Conclusions:

• The DRAL + ADN framework demonstrates significant potential for enhancing deep learning model performance.
• This approach is particularly effective for pathological image datasets with partially mislabeled training data.
• The proposed methods offer a viable solution for improving automated pathological image analysis.