[Paper Review] Detection of Parasitic Eggs from Microscopy Images and the emergence of a new dataset

Perla Mayo, Nantheera Anantrasirichai+3 more|arXiv (Cornell University)|Mar 6, 2022

Digital Imaging for Blood Diseases8 citations

TL;DR

This paper proposes a two-stage deep learning framework for detecting parasitic eggs in low-quality microscopy images using a Pix2Pix GAN for image enhancement followed by Faster R-CNN for object detection. The method achieves improved detection performance—especially in recall and F1-score—by training the detector on GAN-enhanced images, and introduces a new, publicly available dataset to support future research in this domain.

ABSTRACT

Automatic detection of parasitic eggs in microscopy images has the potential to increase the efficiency of human experts whilst also providing an objective assessment. The time saved by such a process would both help ensure a prompt treatment to patients, and off-load excessive work from experts' shoulders. Advances in deep learning inspired us to exploit successful architectures for detection, adapting them to tackle a different domain. We propose a framework that exploits two such state-of-the-art models. Specifically, we demonstrate results produced by both a Generative Adversarial Network (GAN) and Faster-RCNN, for image enhancement and object detection respectively, on microscopy images of varying quality. The use of these techniques yields encouraging results, though further improvements are still needed for certain egg types whose detection still proves challenging. As a result, a new dataset has been created and made publicly available, providing an even wider range of classes and variability.

Motivation & Objective

• To develop an automated, robust method for detecting parasitic eggs in microscopy images, especially under suboptimal image quality conditions common in resource-limited settings.
• To address the challenge of variable image quality due to low-cost equipment, lighting changes, motion blur, and color shifts.
• To improve detection performance by integrating a GAN-based image enhancement step prior to object detection.
• To create and release a new, diverse dataset of parasitic eggs to support future research in automated detection.
• To evaluate the generalization capability of the detection model under distribution shifts caused by image degradation.

Proposed method

• A two-stage framework is employed: first, a Pix2Pix GAN is trained to enhance low-quality microscopy images to high-quality representations.
• The enhanced images are then used to train a Faster R-CNN object detection model, which identifies and localizes parasitic eggs in the images.
• The framework is evaluated using 5-fold cross-validation to ensure robustness and generalization across diverse image conditions.
• Image degradation is simulated during testing using augmentations such as color jittering, brightness/contrast changes, rotation, and flipping to assess robustness.
• The model is tested both on original, degraded, and enhanced test images to compare performance across different input conditions.
• The training and evaluation pipeline is designed to isolate the impact of image enhancement on detection performance, with ablation studies on different enhancement strategies.

Experimental results

Research questions

• RQ1 Can a GAN-based image enhancement step improve the performance of parasitic egg detection in low-quality microscopy images?
• RQ2 How does the detection model perform when tested on images with degradation not seen during training, such as motion blur or lighting variation?
• RQ3 Does training the Faster R-CNN detector on GAN-enhanced images lead to better generalization than training on real low-quality images?
• RQ4 What is the impact of different image enhancement techniques—Pix2Pix vs. CycleGAN—on the final detection accuracy?
• RQ5 Can a unified framework that combines image enhancement and detection in a shared feature space outperform the current two-stage approach?

Key findings

• The Faster R-CNN detector trained on images enhanced by the Pix2Pix GAN achieved the highest precision and recall across all evaluated settings, outperforming models trained on raw or CycleGAN-enhanced images.
• The use of Pix2Pix for enhancement significantly reduced false positives and improved detection sensitivity, particularly for challenging egg types.
• Training the detector on GAN-enhanced images led to better generalization on degraded test images compared to training on real low-quality images.
• The framework that combines image enhancement and detection without retraining the detector on degraded data showed the most consistent performance improvements across all metrics.
• CycleGAN-based enhancement resulted in lower recall, suggesting it may introduce artifacts or misrepresentative features that hinder detection.
• The proposed framework demonstrated robustness to various image corruptions, including color jittering, brightness shifts, and blurring, when the enhancement step was applied before detection.