PI-YOLO: dynamic sparse attention and lightweight convolutional based YOLO for vessel detection in pathological...

Area of Science:

• Computational pathology
• Medical image analysis
• Artificial intelligence in oncology

Background:

• Tumor vessel density is a key prognostic and grading marker.
• Pathologist recognition of vessel density suffers from inter-rater bias.
• Automated detection of vessels in pathology images is challenging due to complex backgrounds and small targets.

Purpose of the Study:

• To develop an enhanced YOLOv7-based network, Pathological Images-YOLO (PI-YOLO), for accurate blood vessel detection in pathology images.
• To improve the quantification of tumor vascularity, aiding in more objective tumor grading and prognosis.
• To address limitations in current automated object detection methods for pathological images.

Main Methods:

• Proposed PI-YOLO, integrating BiFormer attention for global feature extraction and CARAFE upsampling for small target optimization.
• Utilized GSConv module to enhance the ELAN module, improving efficiency and speed.
• Compared PI-YOLO against established object detection models like Faster-RCNN, SSD, RetinaNet, YOLOv5, and YOLOv7.

Main Results:

• PI-YOLO achieved a mean Average Precision (mAP) of 87.48%, outperforming YOLOv7 by 2.83%.
• Demonstrated superior performance on the ICPR 2012 mitotic dataset, achieving an F1 score of 0.8678.
• PI-YOLO showed higher detection accuracy compared to Faster-RCNN, SSD, RetinaNet, and YOLOv5.

Conclusions:

• PI-YOLO offers a significant advancement in automated blood vessel detection in complex pathology images.
• The proposed network enhances accuracy and efficiency, reducing reliance on subjective pathologist interpretation.
• PI-YOLO shows strong potential for improving tumor grading and prognostic value in digital pathology.