A fast and effective detection framework for whole-slide histopathology image analysis

Area of Science:

• Digital Pathology
• Computational Pathology
• Medical Image Analysis

Background:

• The rise of whole-slide imaging in histopathology necessitates advanced computational tools for efficient diagnosis.
• Pathologists' workflow is evolving with the integration of digital pathology and computer-assisted analysis.
• Existing methods require significant computational resources, driving the need for lightweight solutions.

Purpose of the Study:

• To develop a novel, lightweight deep learning framework for automatic tumor detection in whole-slide histopathology images.
• To enhance the accuracy and efficiency of tumor identification in digital pathology slides.
• To create a system that assists pathologists by providing reliable tumor probability heatmaps.

Main Methods:

• Development of a Double Magnification Combination (DMC) classifier, a modified DenseNet-40 with 0.3 million parameters.
• Implementation of an improved adaptive sampling method incorporating superpixel segmentation and a local sampling density heuristic.
• Utilizing a 4-convolutional layer CNN for postprocessing patch-level predictions and linear interpolation for heatmap generation.

Main Results:

• The developed framework achieved an average Area Under the Curve (AUC) of 0.95 for pixel-level tumor detection in the test set.
• The lightweight DMC classifier demonstrated efficient patch-level predictions.
• The integrated system successfully generated tumor probability heatmaps from whole-slide images.

Conclusions:

• The proposed lightweight detection framework offers a promising approach for automatic tumor detection in digital histopathology.
• The system's high performance, validated on challenging datasets, suggests its clinical utility in aiding pathologists.
• This work contributes to the advancement of computer-assisted diagnosis in digital pathology by providing an accurate and efficient tool.