Assessing the quality of whole slide images in cytology from nuclei features

Abstract

Background and objective: Implementation of machine learning and artificial intelligence algorithms into digital pathology laboratories faces several challenges, notably the variation in whole slide image preparation protocols. The diversity of preparation pipelines forces algorithms to be protocol-dependant. Moreover, the error susceptibility of each stage in the preparation process implies a need of quality control tools. To address these challenges, this article introduces a straightforward, interpretable, and computationally efficient quality control module to ensure optimal algorithmic performance.

Methods: The proposed quality control module ensures algorithmic performance by representing an algorithm by a reference whole slide image preparation protocol validated on it. Then, inspired by data description methods, a preparation protocol is represented by nuclei feature distributions, obtained for several whole slide images it has produced. The quality of a preparation protocol is evaluated according to several reference preparation protocols, by comparing their feature distributions with a weighted distance.

Results: Through empirical analysis conducted on seven distinct preparation protocols, we demonstrated that the proposed method build a quality module that clearly discriminates each preparation. Additionally, we showed that this module performs well on more larger and realistic corpus from laboratories routine, detecting quality deviations.

Conclusion: Even if the proposed method necessitates minimal data and few computational resources, we showed that it is interpretable and relevant on realistic corpus from several laboratories' routine. We strongly believe in the necessity of quality control from the algorithmic perspective and hope this kind of approach will be extended to improve quality and reliability of digital pathology whole slide images.

Keywords: Automated quality control; Cytology; Digital pathology; Out-of-distribution detection; Whole slide image analysis.

Fig. 1

Overview of a typical cytological WSI analysis workflow. The analysis module icon represents an analysis with a classification objective.

Fig. 2

Prototypical QC module framework. Modified version of Fig. 13 in Brixtel et al. The ‘Analysis workflow’ refers to Fig. 1. Hatched area indicates where QC is performed in the present work.

Fig. 3

Overview of the dissimilarity measure between two WSI-PPs, w.r.t. to the selection matrix P. WSI-PP representations are obtained via sets of WSIs. These pipelines are then represented via a subset of features depicted by the selection matrix P. For each of these features, the distance function w is performed. Final dissimilarity distance is obtained after analyzing these distances by the aggregator function g.

Fig. 4

Stages to compute reference weights for a WSI-PP (1) with regards to WSI-PP (2) and (3). (a): Reference representation for WSI-PP (1) is computed. (b): Reduction matrix for WSI-PP (1) is obtained from reference representations of WSI-PPs (1), (2), and (3). (c): Distances between WSI-PPs validation (2*) and (3*) (which are pipelines with same characteristics than (2) and (3)) and WSI-PP reference (1) are computed. Reference weight for WSI-PP (1) is defined from these distances. (d): Distance between WSI-PP validation representation (1*) and WSI-PP reference representation (1) is computed. Constraints from Eqs. (3), (4) are checked in order to validate the reference weight.

Fig. 5

Relevant combinations according to $\mathcal{H}$ and $1-\mathcal{E}$. Circled combinations are those selected for ${\mathrm{\Omega }}^{-}$. Combinations comb₂₀, comb₇ and comb₁₈ are those retained for ${\mathrm{\Omega }}^{\star }$.

Fig. 6

Frequency of the parameter values. Each style (color and marker) corresponds to a parameter.

Fig. 7

Evolution of error $\mathcal{E}$ for 100 seeds w.r.t. $n_{\mathrm{nuc}}$ and $n_{\mathrm{wsi}}$.

Fig. 8

Quality measure for each representative pipeline with respect to the batches created for laboratory L. The horizontal dotted black bar is used to separate 2020 and 2021 corpus.

Fig. 9

Quality measure for each representative pipeline with respect to the batches created for laboratory C. The horizontal dotted black bar is used to separate 2019 and 2024 corpus.

Fig. 10

Continuous probability density curve for feature green_min. Curves obtained by fitting a kernel density estimate on the observed feature values.

Fig. 11

Nuclei comparison between laboratory C, pipeline₄ and pipeline₃. Nuclei with the median value for feature green_min have been chosen for pipeline₄ (a), laboratory C pipeline 2019 (b), laboratory C pipeline 2024 (c) and pipeline₃ (d).

Fig. 12

Quality measure for each representative pipeline with respect to the batches created for laboratory M, when considering pipeline 4 and 5 as same.

Fig. 13

Features used for the quality measure of laboratory M w.r.t. to different WSI-PP. Features characterizing colors are represented by the respective color (green, red, or blue), whereas others are represented in black. Quality measure is represented via the dotted line. (For interpretation of the references to color in this figure legend, the reader is referred to the web version of this article.)

Fig. 14

Continuous probability density curve for feature red_max. Curves obtained by fitting a kernel density estimate on the observed feature values.

Fig. 15

Set of nuclei images for laboratory M. Nuclei with the median value for feature red_max have been chosen for 3 batches of laboratory M pipeline: (a) batch 5; (b) batch 26; (c) batch 43.

Fig. 16

Relation between our quality measure and algorithm's confidence.

Fig. A.1

Relevance of combinations, including ${\varphi }_{\mathrm{nuc}}^{\text{debris}}$ parameter, according to $\mathcal{H}$ and $\mathcal{E}$. Circled combinations are those selected for ${\mathrm{\Omega }}^{-}$. Combination comb^★ is selected after Section 4 validation process.

Fig. A.2

Evolution of the error $\mathcal{E}$ for 100 seeds and for debris combinations, w.r.t. $n_{\mathrm{nuc}}$ and $n_{\mathrm{wsi}}$.

Fig. A.3

Quality measure for each pipeline of our metric with respect to the batches created for laboratory M.