Enabling Toxicologic Pathologists Using QuPath, An Open

Source Image Analysis Software Solution

Daniel Rudmann1, Abigail Godbold1, Maureen O’Brien2, Brent Walling1
1 Ashland OH, 2 Frederick MD

1 ABSTRACT 2 METHO DS 4 CO NCLUSIO NS

Anatomic pathology is often a semi-quantitative science and assessing certain microscopic changes in
preclinical studies can be laborious and prone to diagnostic drift. Computer-based image analysis (IA)
could simplify numerous tasks common for the bench pathologist such as setting thresholds and
establishing grading criteria for various tissue changes. Commercial IA solutions are available, but
licenses are typically expensive and may limit use to a single computer or workstation. Historically, open
source software image analysis solutions lacked the sophistication to tackle analyses using whole slide
images (WSI). Recently, QuPath emerged as an alternative open-source platform for IA and aims to
help improve the speed, objectivity, and reproducibility of digital pathology analysis in WSI
https://qupath.github.io/. We developed algorithms in QuPath for 3 common diagnoses/measurements
recorded in preclinical toxicology studies. QuPath was easily downloaded locally on computers,
performed well on a variety of laptops and desktops, and did not require large storage space for
analyses. Training pathologists required a modest time investment and a job aide was easily tailored
from materials produced by QuPath developers. Numerous types of raw image files from different WSI
scanners as well as conventional microscope cameras were dropped quickly into QuPath for
analysis. The region of analysis (ROA) was simple to define using practical annotation
tools. Algorithms were developed to evaluate the 3 diagnoses and preliminary data suggested they
increased the diagnostic confidence and efficiency for the pathologists. We conclude that QuPath
appears to have good potential as a tool for bench toxicologic pathologists.
6 different quantitative challenges were evaluated in this pilot (Table 1) • QuPath was easily downloaded locally on computers, performed well on a variety of laptops and
desktops, and did not require large storage space for analyses when external drives or Cloud servers
• Hepatocellular hypertrophy- size/volume (rat) were used
• Thyroid follicular cell hypertrophy- size/volume (rat)
• Tissue cross reactivity (TCR)- positive cell detection (human) • Numerous types of raw image files from different WSI scanners as well as conventional microscope
• Bone marrow cellularity- cell number (rat) cameras were dropped quickly into QuPath for analysis
• Delayed type hypersensitivity- cell number (monkey)
• Cell proliferation Ki67 immunohistochemistry- positive cell detection (human) • The region of analysis (ROA) was simple to define using practical annotation tools. The “Polygon”
tool allowed tracing of a ROA. The “Wand” tool enabled faster annotation and supported exclusions
General QuPath methods (done for all challenges)
• Slides were manually cleaned and scanned on a Leica AT or AT2 at 20 or 40x • Default cell detection settings were good for preliminary nuclear detection and cell counts
• QuPath v 0.1.2 was downloaded at https://qupath.github.io/ using “without administrative rights”
• WSS files (svs) for analysis were directly “dragged” into QuPath for analysis (Fig 1A) • For assessment of cell numbers, smoothing, nuclear area, and DAB/hematoxylin thresholds were
• The polygonal or wand tool was used to annotate images (Fig 1B) useful in targeting cell populations (lymphocytes, hepatocytes, bone marrow cells, thyroid follicular
• For slides involving detection of DAB chromogen (TCR, Ki67): AnalyzeCell AnalysisPositive Cell cells) while excluding other cells or artifact not part of the ROA
Detection was used followed (Fig 1C)
• AnalyzeCell AnalysisCell Detection was used at default settings for other non-IHC IA • For defining cells in support of cell volume measurements, modifying cell expansion parameters and
• Quality review was done by a pathologist to verify cell detection was acceptable based on “challenge” applying nuclear distance measurements were helpful
• If quality review failed, detection settings were modified empirically
• Algorithm settings then applied to additional tissues (Fig 1D) or animals • QuPath is still in an early release version but based on this preliminary work and the clear
engagement of the developer, it appears to have good potential for utility in toxicologic pathology

Table 1: Study Type/Diagnoses for QuPath IA Analysis and Potential Figure 2: Hepatocellular Hypertrophy- Nuclear Selection Figure 3: Hepatocellular Hypertrophy- Nuclear to

3 Nuclear Distance Analysis

Return on Investment
RESULTS Study Type/Diagnosis
Figure 1A-D: General Approach and Effort for IA Set Up in QuPath (Approximate Time Hepatocellular Hypertrophy
for Each Step in Arrow)
Thyroid Follicular Cell
A: Dropping Analysis File B: Annotation Hypertrophy
Tissue Cross Reactivity
Bone Marrow Cellularity
Delayed Type Hypersensitivity
Endpoint(s)
Cell Size, Nuclear-Nuclear
Distance
Cell Size, Nuclear-Nuclear
Distance
Positive Cell Number (1-3+)
Total Cell Number or Cell
Number/Area
Total Cell Number or Cell
Number/Area
Selection of hepatocytes by adjusting cell/nuclear parameters in QuPath
Potential ROI
Speed, grading, and
consistency
Distance
Speed, grading, and threshold
consistency
Threshold setting, speed,
and consistency
Speed, grading, and Initial cell/nuclear detection Optimized selection of hepatocytes Minimizing nuclear-nuclear distance biased when crossing large areas/sinusoids (*).
consistency By setting maximum distance threshold, nuclear-nuclear distance better estimated.
Figure 4: Thyroid Follicular Cell Analysis Figure 5: Thresholding Approach- TCR Studies
Speed, grading, and Follicular Precipitates Follicular Cell Epithelial Gaps Optimized Image
Negative 1+ 1-2+
consistency

< 5 sec < 10 sec Cell Proliferation Positive Cell Number or Threshold setting, speed,
Positive Cell Number/Area and consistency

Table 2: IA Solutions- Challenges, Key Parameters, and Probability of Default Settings Increasing Hematoxylin Threshold Increasing Cell Expansion
10 sec- 2 min

Technical Success (PTS)

Study Challenge(s) Key Tool or Parameter(s) PTS
Type/Diagnosis
Hepatocellular Cell border, nuclear Cell expansion, nuclear area, Medium
Hypertrophy distance (Figs. 2-3) intensity and nuclear distance Adjusting thresholds to the eye of the Pathologist to establish parameters-applying
same parameters from slide to slide prevented diagnostic drift
Thyroid Cell border definition Wand annotation tool, cell Medium
Follicular Cell (Fig 4) expansion, intensity threshold Figure 6: Bone Marrow Cellularity- Nuclear Selection Figure 7: DTH Studies- Counting Lymphocytes Figure 8: Cell Proliferation- Ki67 Tonsillar Enumeration
Hypertrophy parameters
<2 min/file A C
Tissue Cross Thresholds (Fig 5) Intensity threshold parameters High
Reactivity (DAB OD)
Bone Marrow Nuclear Nuclear area + intensity Medium
Cellularity differentiation (Fig 6) threshold (hematoxylin OD) B D
Cell # 17K 24K 36K
Delayed Type Accuracy (Fig 7) Nuclear area + intensity High
Hypersensitivity threshold (hematoxylin OD)
D: Quality Review by Pathologist and Evaluating C: Positive Cell Detection
Cell Cell border definition Intensity threshold (DAB OD), High
Additional Tissues/Animals Proliferation (Fig 7-8) smoothing, nuclear area Default settings miss some Reduction of minimum nuclear
RBC islands area by 50% detected RBC islands
Effects of adjusting nuclear area and threshold to increase accuracy of Initial run (A, B) highlighting positive (red) and negative (blue) nuclear detection and
lymphocyte selection (positive cells red) ; pathologists quality review arrows of nuclei that escaped detection. Optimized (C,D) using smoothing parameters,
indicated middle settings were most accurate (24K cell count) altered DAB thresholds, and adjusted nuclear size detection