[Paper Review] Deep learning-enabled image quality control in tomographic reconstruction: Robust optical diffraction tomography

DongHun Ryu, YoungJu Jo+7 more|Figshare|Mar 5, 2019

Digital Holography and Microscopy35 references4 citations

TL;DR

This paper proposes a deep learning-based quality control system for optical diffraction tomography that automates the detection of defective holographic data, using a convolutional neural network trained on a large annotated dataset of clean and noisy optical field images. The model achieves over 90% test accuracy, outperforming both non-expert visual inspection and rule-based methods, and significantly enhances tomographic reconstruction quality by filtering out low-quality raw data.

ABSTRACT

In tomographic reconstruction, the image quality of the reconstructed images can be significantly degraded by defects in the measured two-dimensional (2D) raw image data. Despite the importance of screening defective 2D images for robust tomographic reconstruction, manual inspection and rule-based automation suffer from low-throughput and insufficient accuracy, respectively. Here, we present deep learning-enabled quality control for holographic data to produce robust and high-throughput optical diffraction tomography (ODT). The key idea is to distill the knowledge of an expert into a deep convolutional neural network. We built an extensive database of optical field images with clean/noisy annotations, and then trained a binary classification network based upon the data. The trained network outperformed visual inspection by non-expert users and a widely used rule-based algorithm, with > 90% test accuracy. Subsequently, we confirmed that the superior screening performance significantly improved the tomogram quality. To further confirm the trained model's performance and generalizability, we evaluated it on unseen biological cell data obtained with a setup that was not used to generate the training dataset. Lastly, we interpreted the trained model using various visualization techniques that provided the saliency map underlying each model inference.

Motivation & Objective

• To address the challenge of low-throughput and inaccurate manual or rule-based screening of defective 2D raw image data in tomographic reconstruction.
• To develop an automated, scalable solution for image quality control in optical diffraction tomography (ODT) using deep learning.
• To improve the robustness and quality of reconstructed tomograms by eliminating defective input data before reconstruction.
• To ensure generalizability of the trained model across different biological cell samples and imaging setups not included in the training data.
• To interpret the model's decision-making process using explainability techniques such as saliency maps.

Proposed method

• A large-scale database of optical field images was constructed with clean and noisy annotations to train a deep learning model.
• A binary classification convolutional neural network (CNN) was trained to distinguish between clean and defective holographic data.
• Knowledge distillation from expert annotations was used to enhance the model's performance and generalization.
• The trained model was applied to screen raw 2D holographic data prior to tomographic reconstruction.
• Saliency maps and visualization techniques were used to interpret the model’s predictions and identify key features influencing classification.
• The method was validated on unseen biological cell data from a different imaging setup than used in training.

Experimental results

Research questions

• RQ1 Can a deep learning model outperform non-expert human inspection in identifying defective holographic data for optical diffraction tomography?
• RQ2 Can a deep learning-based quality control system improve the accuracy and robustness of tomographic reconstructions?
• RQ3 How generalizable is the trained model to new biological samples and imaging conditions not present in the training data?
• RQ4 What features does the model rely on when classifying image quality, and can these be interpreted meaningfully?
• RQ5 Can knowledge distillation from expert annotations enhance the performance of a deep learning model for image quality control?

Key findings

• The deep learning model achieved over 90% test accuracy in classifying clean versus defective holographic images, surpassing both non-expert visual inspection and a widely used rule-based algorithm.
• The use of the trained model for pre-screening raw data led to a significant improvement in the quality of reconstructed tomograms.
• The model generalized well to unseen biological cell data collected with a different imaging setup than used during training.
• Saliency map analysis revealed that the model focused on relevant structural features in the holograms, indicating meaningful and interpretable decision-making.
• The model’s performance was robust across diverse biological samples, demonstrating its potential for real-world deployment in high-throughput ODT systems.
• The integration of deep learning for image quality control reduced reliance on manual inspection and enabled scalable, automated data validation in tomographic reconstruction pipelines.