Hierarchical Object Detection for Small Lesions in Whole Slide Images
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current methods for microscopic anatomic pathology diagnosis, particularly in placental histopathology, are time-consuming and resource-intensive, with limited automation and high false positive rates, especially in detecting small lesions like decidual vasculopathy (DV) in placentas, which are crucial for predicting pregnancy complications like preeclampsia.
Innovation Solution
A hierarchical deep learning framework that includes object detection and classification stages, using RetinaNet for localization and Xception for classification, combined with patient metadata, to accurately identify DV lesions in whole slide images (WSIs) and aggregate latent feature representations for improved diagnostic accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If deep learning is used to detect small lesions like decidual vasculopathy in whole slide images, then detection accuracy is improved, but false positive rates increase due to the large number of patches analyzed
Solution Approach 1:
The system divides the whole slide image into multiple patches and further segments each patch into smaller regions of interest. This segmentation allows the deep learning model to focus on specific areas, reducing false positives while maintaining detection accuracy across the entire large image.
Solution Approach 2:
The patent applies different processing strategies to different regions. Small lesions in specific patches are identified and analyzed with higher precision, while other regions are processed more efficiently. This local quality approach ensures high detection accuracy where needed without compromising overall reliability.
2Measurement precision
If manual microscopic inspection is performed on all tissue sections, then diagnostic accuracy is maintained, but time consumption increases significantly
Solution Approach 1:
The patent replaces manual microscopic inspection with an automated deep learning system. The mechanical process of manual scanning and analysis is substituted by computational algorithms that can process entire whole slide images rapidly, maintaining diagnostic accuracy while dramatically reducing time consumption.
Solution Approach 2:
The system performs self-service analysis by automatically detecting, localizing, and characterizing lesions without requiring continuous human intervention. The deep learning model independently processes images, identifies patterns, and provides diagnostic assistance, freeing pathologists from routine inspection tasks.
3Productivity
If automated deep learning analysis is implemented, then productivity increases, but device complexity increases due to multiple processing stages
Solution Approach 1:
The complex analysis process is segmented into distinct stages: whole slide image processing, patch extraction, region of interest identification, and lesion characterization. Each stage is handled by a specialized module, making the overall complex system more manageable and maintainable while achieving high productivity.
Solution Approach 2:
The patent implements a nested structure where smaller analysis units (patches and regions of interest) are contained within larger structures (whole slide images). This nesting allows the system to handle complex analysis by recursively processing smaller units, improving productivity without proportionally increasing operational complexity.
Data Source
AI summary
A hierarchical deep-learning object detection framework provides a method for identifying objects of interest in high-resolution, high pixel count images, wherein the objects of interest comprise a relatively a small pixel count when compared to the overall image. The method uses first deep-learning model to analyze the high pixel count images, in whole or as a patchwork, at a lower resolution to identify objects, and a second deep-learning model to analyze the objects at a higher resolution to classify the objects.