Efficient Tissue Detection in Whole-Slide Images Using Classical and Hybrid Methods: Benchmark on TCGA Cancer...

Area of Science:

• Digital Pathology
• Computational Biology
• Artificial Intelligence in Medicine

Background:

• Whole-slide images (WSIs) are vital for digital pathology, enabling large-scale cancer pattern analysis.
• WSI artifacts and non-tissue regions hinder AI processing, increasing computational costs and errors.
• Accurate tissue detection is crucial for WSI pipelines, but deep learning methods demand extensive manual annotations.

Purpose of the Study:

• To benchmark thumbnail-level tissue detection methods for whole-slide images.
• To evaluate the accuracy, speed, and efficiency of annotation-free versus supervised approaches.
• To introduce and validate the novel Double-Pass hybrid method for WSI preprocessing.

Main Methods:

• Benchmarking Otsu's thresholding, K-Means clustering, GrandQC's UNet++, and the annotation-free Double-Pass hybrid method.
• Utilizing 3322 TCGA whole-slide images across nine cancer cohorts for evaluation.
• Assessing performance based on mean Intersection over Union (mIoU), processing speed, and computational efficiency on CPU.

Main Results:

• The Double-Pass method achieved a high mIoU of 0.826, comparable to the deep learning UNet++ model (0.871).
• Double-Pass processed slides significantly faster (0.203 s/slide) than UNet++ (2.431 s/slide) on a CPU.
• The annotation-free, CPU-optimized Double-Pass method demonstrated superior efficiency for scalable tissue detection.

Conclusions:

• The annotation-free Double-Pass pipeline offers a scalable solution to computational bottlenecks in WSI preprocessing.
• This method facilitates high-throughput analysis, enabling faster and more cost-effective AI integration in pathology.
• Double-Pass presents a novel, fast, and robust alternative to supervised methods for tissue detection in digital pathology.