Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Gaze-DETR: Using Expert Gaze to Reduce False Positives in Vulvovaginal Candidiasis Screening

Yan Kong, Sheng Wang, Jiangdong Cai, Zihao Zhao, Zhenrong Shen, Yonghao Li, Manman Fei, Qian Wang

TL;DR

This work tackles false positives in vulvovaginal candidiasis detection from Thin-prep cytologic test images by leveraging expert gaze data. It introduces Gaze-DETR, a training framework with a gaze-guided warm-up and a gaze-guided rectification stage that uses gaze-based signals during training while keeping inference gaze-free. Key contributions include the definition of 'gaze only' boxes, a universal warm-up strategy adaptable to CNN and DETR-like detectors, and a rectification mechanism that concentrates training queries on confounding regions. Empirical results on an in-house dataset show improved precision and generalization over state-of-the-art detectors, with broad compatibility across backbones and detection paradigms, highlighting the potential of human expert gaze to enhance medical image analysis.

Abstract

Accurate detection of vulvovaginal candidiasis is critical for women's health, yet its sparse distribution and visually ambiguous characteristics pose significant challenges for accurate identification by pathologists and neural networks alike. Our eye-tracking data reveals that areas garnering sustained attention - yet not marked by experts after deliberation - are often aligned with false positives of neural networks. Leveraging this finding, we introduce Gaze-DETR, a pioneering method that integrates gaze data to enhance neural network precision by diminishing false positives. Gaze-DETR incorporates a universal gaze-guided warm-up protocol applicable across various detection methods and a gaze-guided rectification strategy specifically designed for DETR-based models. Our comprehensive tests confirm that Gaze-DETR surpasses existing leading methods, showcasing remarkable improvements in detection accuracy and generalizability.

Gaze-DETR: Using Expert Gaze to Reduce False Positives in Vulvovaginal Candidiasis Screening

TL;DR

This work tackles false positives in vulvovaginal candidiasis detection from Thin-prep cytologic test images by leveraging expert gaze data. It introduces Gaze-DETR, a training framework with a gaze-guided warm-up and a gaze-guided rectification stage that uses gaze-based signals during training while keeping inference gaze-free. Key contributions include the definition of 'gaze only' boxes, a universal warm-up strategy adaptable to CNN and DETR-like detectors, and a rectification mechanism that concentrates training queries on confounding regions. Empirical results on an in-house dataset show improved precision and generalization over state-of-the-art detectors, with broad compatibility across backbones and detection paradigms, highlighting the potential of human expert gaze to enhance medical image analysis.

Abstract

Accurate detection of vulvovaginal candidiasis is critical for women's health, yet its sparse distribution and visually ambiguous characteristics pose significant challenges for accurate identification by pathologists and neural networks alike. Our eye-tracking data reveals that areas garnering sustained attention - yet not marked by experts after deliberation - are often aligned with false positives of neural networks. Leveraging this finding, we introduce Gaze-DETR, a pioneering method that integrates gaze data to enhance neural network precision by diminishing false positives. Gaze-DETR incorporates a universal gaze-guided warm-up protocol applicable across various detection methods and a gaze-guided rectification strategy specifically designed for DETR-based models. Our comprehensive tests confirm that Gaze-DETR surpasses existing leading methods, showcasing remarkable improvements in detection accuracy and generalizability.
Paper Structure (11 sections, 4 figures, 3 tables)

This paper contains 11 sections, 4 figures, 3 tables.

Table of Contents

  1. Introduction
  2. Method
  3. Gaze Processing
  4. Gaze-Guided Warm-up
  5. Gaze-Guided Rectification
  6. Experimental Results
  7. Data and Experimental Settings
  8. Comparison with the State-of-the-Art
  9. Compatibility of our method
  10. Ablation Study
  11. Conclusion

Figures (4)

  • Figure 1: (A) TCT image samples displaying false positives (red) versus true positives (green), and corresponding gaze heatmap. (B) Gaze tracking during labeling: An expert's gaze (yellow) and mouse-labeled regions on a TCT image. Regions where experts carefully review without annotating are usually confounding area.
  • Figure 2: Overview of the Gaze-DETR. (A) Illustration of Gaze Processing, which derives gaze points to 'gaze only' boxes. (B) Pipeline of the Gaze-DETR model, which integrates the gaze guidance into the DETR framework for candida detection.
  • Figure 3: Illustration to compare ‘no object’ queries of standard DETR against Gaze-DETR. Gaze-DETR effectively leverages 'gaze only' boxes to concentrate 'no object' queries on confounding regions, compared to the nearly random distribution of 'no object' queries of DETR.
  • Figure 4: Comparison of qualitative detection results. Our method significantly reduces false positive errors in areas that are carefully reviewed by experts, which are often prone to misclassifications by both RetinaNet and DN-DETR.