Focus Quality Assessment of High-Throughput Whole Slide Imaging in Digital Pathology

Mahdi S. Hosseini, Yueyang Zhang, Lyndon Chan, Konstantinos N. Plataniotis, Jasper A. Z. Brawley-Hayes, Savvas DamaskinosIEEE TMI 2018 · OpenReview

TL;DR

This paper introduces FQPath, a novel no-reference focus quality assessment (NR-FQA) metric specifically designed for digital pathology whole slide images (WSIs). FQPath leverages a synthesized human visual system (HVS)-like kernel to accurately quantify image focus at both the patch and slide level, offering a computationally efficient solution for automated quality control in digital pathology workflows.

What Problem Does This Paper Address?

The widespread adoption of digital pathology, which involves scanning glass tissue slides into high-resolution digital images (WSIs), is hampered by the need for robust automated quality control. A critical aspect of this quality control is ensuring the focus of the scanned images. Inaccurate focusing by WSI scanners can render images unusable, leading to the need for re-scanning, which is time-consuming and costly, especially in high-throughput environments. Existing NR-FQA methods often struggle with the unique characteristics of WSIs, such as their large size, diverse tissue types, and the specific nature of blur (primarily out-of-focus rather than motion blur). Furthermore, many existing methods are computationally intensive, making them impractical for real-time clinical applications. This paper aims to bridge this gap by developing an accurate, efficient, and automated NR-FQA metric tailored for digital pathology WSIs.

Methods

The core of FQPath lies in its novel approach to synthesizing a Human Visual System (HVS)-like kernel. This kernel is designed to mimic the HVS's ability to boost high-frequency information, thereby counteracting the high-frequency fall-off caused by the out-of-focus optics of the WSI scanner.

  1. PSF Modeling: The paper first models the point spread function (PSF) of the out-of-focus scanner optics using the Born & Wolf model.
  2. HVS-M Kernel Synthesis: The inverse of this PSF is then approximated by a linear combination of even-derivative filters from the MaxPol library. This synthesized kernel, termed HVS-M, is designed to equalize the frequency spectrum of the image, making it more sensitive to focus-related features.
  3. Patch-Level Feature Extraction: The HVS-M kernel is convolved with grayscale WSI patches. Positive features from the horizontal and vertical convolutions are activated, vectorized, and then processed to extract dominant focus-relevant features.
  4. Focus Quality Score: A focus quality score is derived from the energy of these features using an $l_{1/2}$-norm and a central moment calculation. This score is inversely proportional to focus quality (lower score = better focus).
  5. Slide-Level Heatmap Generation: For whole slide analysis, FQPath is applied to individual patches. To linearize the focus quality gradation for better visual perception, an inverse Gaussian projection is applied to the patch scores. These projected scores are then used to generate a local focus quality heatmap for the entire slide.

Datasets

  • FocusPath Database: The paper utilizes a subset of the FocusPath database for training and testing the inverse Gaussian projection. This database contains 540 training profiles and 90 testing profiles, each with ground-truth z-levels.
  • 200-WSI Scan Database: A new database of 200 WSIs, scanned at 40X magnification (0.25µm/pixel) without compression, was collected for evaluating the slide-level focus quality assessment. These slides are H&E stained and cover various tissue types and diagnoses.
  • Subjective Scoring Protocol: For validation, a subjective scoring protocol was developed for the 200-WSI database, where human assessors graded the focus quality of each slide.

Key Insights and Ideas

  • HVS Mimicry for Focus Assessment: The core innovation is the use of an HVS-inspired kernel to counteract the specific blur introduced by WSI scanners. This approach moves beyond generic blur assessment by tailoring the metric to the optical characteristics of the imaging system.
  • Patch-to-Slide Coherence: The ability to generate a localized focus quality heatmap from patch-level scores is crucial. The inverse Gaussian projection is a clever step to ensure that the visual representation of focus quality on the heatmap is intuitive and not skewed by non-linear score distributions.
  • Computational Efficiency as a Design Goal: The authors explicitly prioritize computational efficiency, recognizing its critical role in clinical adoption. Their method is designed to be fast enough for real-time applications, a significant advantage over many existing complex NR-FQA techniques.
  • Focus Quality as a Proxy for Scan Quality: The paper demonstrates that focus quality is a strong indicator of overall scan quality, making FQPath a valuable tool for automated quality control that can help clinicians and pathologists.

Key Results

  • Patch-Level Accuracy: FQPath achieved the highest overall performance across multiple accuracy measures (PLCC, SRCC, KRCC, RMSE) when compared to ten state-of-the-art NR-FQA metrics on the FocusPath database. It showed significantly better accuracy than 4 other metrics and comparable performance to others.
  • Computational Speed: FQPath demonstrates excellent computational efficiency. For a 1024x1024 patch, it takes only 0.06626 seconds, significantly faster than methods like MIFQ (74 times slower).
  • Slide-Level Heatmap Utility: The generated focus quality heatmaps visually highlight regions of good and poor focus, aiding in the identification of problematic scans.
  • Correlation with Subjective Scores: The objective slide-level acceptance ratios derived from FQPath heatmaps showed a linear relationship with subjective acceptance ratios, indicating the feasibility of using FQPath for automated slide-level quality control.

Limitations and Future Work

  • Parameter Tuning: The FQPath metric requires tuning of parameters ($z^*$, $w^*$, $m^*$). While the paper describes a grid search for optimal parameters, this tuning process might need to be revisited for different scanner types or imaging conditions.
  • Generalization to Other Modalities: While the authors suggest the metric's potential for other imaging modalities (e.g., fluorescence microscopy), its performance on these would require further validation.
  • Future Work: The authors suggest that the metric could be extended to analyze the relevance of out-of-focus regions for automated diagnosis tools, opening avenues for more advanced applications in computational pathology. They also highlight the potential for using FQPath in other medical imaging applications for QC control.

Summary updated

Abstract

One of the challenges facing the adoption of digital pathology workflows for clinical use is the need for automated quality control. As the scanners sometimes determine focus inaccurately, the resultant image blur deteriorates the scanned slide to the point of being unusable. Also, the scanned slide images tend to be extremely large when scanned at greater or equal 20X image resolution. Hence, for digital pathology to be clinically useful, it is necessary to use computational tools to quickly and accurately quantify the image focus quality and determine whether an image needs to be re-scanned. We propose a no-reference focus quality assessment metric specifically for digital pathology images, that operates by using a sum of even-derivative filter bases to synthesize a human visual system-like kernel, which is modeled as the inverse of the lens' point spread function. This kernel is then applied to a digital pathology image to modify high-frequency image information deteriorated by the scanner's optics and quantify the focus quality at the patch level. We show in several experiments that our method correlates better with ground-truth $z$-level data than other methods, and is more computationally efficient. We also extend our method to generate a local slide-level focus quality heatmap, which can be used for automated slide quality control, and demonstrate the utility of our method for clinical scan quality control by comparison with subjective slide quality scores.
BibTeX
@article{hosseini2018focus,
  author = {Mahdi S. Hosseini and Yueyang Zhang and Lyndon Chan and Konstantinos N. Plataniotis and Jasper A. Z. Brawley-Hayes and Savvas Damaskinos},
  title = {{Focus Quality Assessment of High-Throughput Whole Slide Imaging in Digital Pathology}},
  journal = {IEEE Trans. Medical Imaging},
  year = {2018},
  url = {https://openreview.net/forum?id=5eunEqlu5ZWe},
}
Figure