AI SOFTWARE DEVELOPMENT
Pathology Visions Poster Presentation
Utilizing artificial intelligence-assisted heuristic scanning to enhancing efficiency of urine cytology slides scanning
Resources and News
Educational resources and news from AIxMed to keep you informed
and up-to-date on our AI empowered digital cytology solutions.
and up-to-date on our AI empowered digital cytology solutions.
.png)
Thank you! Your submission has been received!
Oops! Something went wrong while submitting the form.
AIxURO URINE CYTOLOGY CLINICAL STUDIES
May 2025
ICC 2025 Oral Presentation
AIxURO: Enhancing Diagnostic Accuracy in Bladder Caner Detection with AI-Assisted Urine Cytology
ICC 2025 Oral Presentation
AIxURO: Enhancing Diagnostic Accuracy in Bladder Caner Detection with AI-Assisted Urine Cytology
STUDY DESIGN:
- Retrospective study using 200 ThinPrep urine cytology slides.
- Ground truth: NHGUC (n=100), AUC (n=35), SHGUC (n=32), HGUC (n=33).
- Slides digitized using: (1) Mikroscan SLxCyto( Mikroscan-WSIs); (2) Customized Huron scanner (Huron-WSIs).
- WSIs were analyzed using AIxURO, a disease-specific AI algorithm for bladder cancer.
- Three-arm blinded review by 1 cytopathologist and 2 cytologists, with 2-week washout periods:
- Arm 1: Conventional microscopy
- Arm 2: AIxURO with Mikroscan-WSIs
- Arm 3: AIxURO with Huron-WSIs
- Performance Metrics:
- Comparison of study diagnosis with urine cytology ground truth diagnosis (n=200) for 3 arms, using following threshold: AUC+(AUC, SHGUC and HGUC) cases as positive; NHGUC cases as negative.
- Assessment of agreement between study diagnoses and surgical pathology-confirmed bladder cancer (n=72) among the three arms.
- Evaluation of bladder cancer prediction in hematuria-indicated cases (n=16) across the three arms.
RESULTS:
- Urine cytology ground truth cases: Arm 2 and Arm3 exhibited higher sensitivity (85.0% and 88.3%) compared to Arm 1 (79.3%) with comparable accuracy across all arms (Arm 1: 86.8%, Arm 2: 85.3%, Arm 3: 85.3%).
- Biopsy-confirmed bladder cancer cases: Arm 2 and Arm 3 showed improved sensitivity (92.0% and 93.2%) and accuracy (82.9% and 82.4%) compared to Arm 1 (84.6% sensitivity, 78.2% accuracy).
- Hematuria cases: Arms 2 and 3 achieved superior sensitivity (96.7% and 100.0%) and accuracy (89.6% and 91.7%) compared to Arm 1 (90.0% sensitivity, 85.4% accuracy).
CONCLUSIONS:
AIxURO consistently enhanced diagnostic sensitivity and maintained accuracy across multiple scanners, outperforming microscopy in detecting bladder cancer. It demonstrated strong predictive performance in high-risk subgroups, supporting its utility in real-world clinical workflows and digital cytology integration.
AI SOFTWARE DEVELOPMENT
May 2025
ICC 2025 - Oral Presentation
Does AI performance in Thyroid Cancer Diagnosis Differ Between Single-Layer and Multi-Z-Stacked Imaging?
ICC 2025 - Oral Presentation
Does AI performance in Thyroid Cancer Diagnosis Differ Between Single-Layer and Multi-Z-Stacked Imaging?
STUDY DESIGN:
- 71 ThinPrep thyroid FNAC slides selected by consensus cytology diagnosis
- Cases included: TBS-II (n=35), TBS-IV (n=6), TBS-VI (n=30), confirmed by biopsy
- Slides digitized using 3DHistech scanner to create paired whole-slide images:
- S-WSI: single-layer WSI
- 7-WSI: seven-layer Z-stacked WSI
- AI-assisted review used an AIxTHY model to detect cancer cells and guide interpretation
- Three cytologists independently reviewed both S-WSI and 7-WSI sets
- Total of 213 reads (71 cases × 3 reviewers) analyzed
- Two diagnostic threshold:
- Threshold 1: Positive = TBS-V/VI; Negative = TBS-II
- Threshold 2: Positive = TBS-IV/V/VI; Negative = TBS-II
- Outcomes measured:
- Binary diagnostic sensitivity and specificity (under both thresholds)
- Agreement with consensus diagnoses
- Interobserver agreement (Cohen’s κ)
RESULTS:
- Threshold 1 (TBS-V/VI vs. TBS-II):
- Sensitivity: 7-WSI 88.9% vs. S-WSI 81.1%
- Specificity: 7-WSI 94.3% vs. S-WSI 96.2% (not significant)
- Threshold 2 (TBS-IV/V/VI vs. TBS-II):
- Sensitivity: 7-WSI 86.1% vs. S-WSI 80.6% (p < 0.05)
- Specificity: 7-WSI (%): All cases: 7-WSI 88.6% vs. S-WSI 91.4%
- Consensus Agreement (%): All cases: 7-WSI 74.6%vs. S-WSI 61.5%
- TBS-VI: 65.6% vs. 36.7%
- TBS-IV: 44.4% vs. 27.8%
- TBS-II: 87.6% vs. 88.6%
- Interobserver Agreement (Cohen’s κ): Overall: 0.366 vs. 0.351
- TBS-IV: 7-WSI 0.189 vs. S-WSI 0.075
- TBS-VI: 0.420 vs. 0.319
- TBS-II: 0.549 vs. 0.604
CONCLUSIONS:
Seven-layer Z-stacked WSIs (7-WSI) enhanced AI-assisted thyroid cancer diagnosis and improved interobserver agreement among cytologists, especially in indeterminate (TBS-IV) and malignant (TBS-VI) cases, supporting their clinical value over single-layer WSIs.
THE PARIS SYSTEM FOR URINE CYTOLOGY NUCLEAR TO CYTOPLASMIC RATIO
May 2025
Academic Article: Cancer Cytopathology
Does Artificial Intelligence Redefine Nuclear-to-Cytoplasmic Ratio Threshold for Diagnosing High-Grade Urothelial Carcinoma?
Academic Article: Cancer Cytopathology
Does Artificial Intelligence Redefine Nuclear-to-Cytoplasmic Ratio Threshold for Diagnosing High-Grade Urothelial Carcinoma?
STUDY DESIGN:
- 106 urine cytology slides (from 46 lower and 60 upper urinary tract HGUC/CIS cases) digitized and analyzed using AIxURO, an AI-based WSI tool.
- AI quantified atypical and suspicious urothelial cells, their nuclear-to-cytoplasmic(N/C) ratios, and nuclear areas.
- Morphologic data were correlated with biopsy-confirmed HGUC/CIS diagnoses. Statistical comparisons were conducted using Kruskal–Wallis tests.
RESULTS:
- Suspicious vs Atypical Cells (Total):
- Fewer suspicious cells detected (median 20.5 vs 242.0, p<.001)
- Higher median N/C ratio: 0.66 vs 0.58 (p<.001)
- Larger nuclear area: 102.3 µm² vs 85.7 µm² (p<.001)
- By Cytology Category (AUC, SHGUC, HGUC):
- Suspicious cells consistently had higher N/C ratios and nuclear areas than atypical cells.
- Median N/C ratio of suspicious cells: 0.66~0.67 and atypical cells: 0.58~0.59 across groups
- Median nuclear areas of suspicious cells: AUC 108.9 µm², SHGUC 99.2 µm², HGUC 101.6 µm² and atypical cells: AUC 88.5 µm², SHGUC 86.8 µm², HGUC 83.5 µm²
- Upper vs Lower Tract:
- No significant difference in N/C ratios
- Nuclear areas were smaller in UUT vs LUT cases (e.g., suspicious: 98.0 µm² vs 108.0 µm², p<.001)
- No significant difference in N/C ratios
- Biopsy Correlation:
- CIS had the largest nuclear areas among suspicious (116.3 µm²) and atypical (101.5 µm²) cells
- Cell number and N/C ratio did not differ significantly across CIS, CIS-HGUC, and HGUC biopsy categories
CONCLUSIONS:
AIxURO provides objective measurement of N/C ratio and nuclear area in urine cytology, challenging the current TPS threshold (>0.7) for diagnosing HGUC. The study suggests that a revised N/C ratio cutoff of 0.66 may be more appropriate for SHGUC/HGUC categorization using AI. Findings also support the use of consistent thresholds across upper and lower urinary tract cases. Nuclear area offers additional discriminative value, particularly for differentiating CIS from HGUC.
October 2023
Utilizing artificial intelligence-assisted heuristic scanning to enhancing efficiency of urine cytology slides scanning
DPA PathVysions 2023 Poster Presentation / Abstract (J Path Inform 2023) Liu TJ, Hang JF, Ou YC et al. TVGHTE and Tung’s Taichung MetroHospital
STUDY DESIGN:
- 52 urine cytology slides from bladder cancer patients (24 cytospin, 16 ThinPrep, 12 Cytorich) were digitally scanned with Leica AT2 using conventional Z-stacking (Z = 0 to Z = +10 layers above and below at 1 µm intervals) and compared with using a heuristic scanning simulation method with AI to determine regions of interest (ROI) for scanning, whereby the software determines the ideal number of layers to scan, from 3 layers (Z = 0+1) to 21 layers (Z = +10).
- Performance Metrics: Total number of suspicious cells; coverage rate (ratio of suspicious cells in single vs multiple Z-layers to the total suspicious cells in 21 Z-layers); scanning time (seconds); and image file size for storage (in gigabytes).
RESULTS:
- Heuristic scanning was comparable to Z-stacking with similar average numbers of suspicious cell coverage rates (79.3% cytospin; 85.9% TP; 78.3% CytoRich) to those of Z-stacking (81.9% in 5 layers for cytospin, 87.1% in 9 layers for TP, 82.9% in 7 layers for CytoRich).
- Heuristic scanning significantly reduced scanning time (65-72%) and image file size (by 46-64%).
- Heuristic scanning in low cellularity slides was more accurate (7/7 correct diagnoses) than using single Z-layer WSI (4/7 correct diagnoses)
CONCLUSIONS:
Heuristic scanning is an effective alternative to conventional Z-stacking to identify suspicious cells on urine cytology slides, with the advantage of reducing scanning time and image file size.
August 2023
SOFTWARE OPTIMIZATION & VALIDATION STUDIES
Japanese Society of Digital Pathology (JSDP) 21st Annual Meeting
Detection of urothelial carcinoma in upper urinary tract cytology using a digital artificial intelligence-assisted tool
Detection of urothelial carcinoma in upper urinary tract cytology using a digital artificial intelligence-assisted tool
2023 Japanese Society of Digital Pathology Oral Platform/ Abstract Liu TJ, Yang WL, Hsu SW, Yeh CH, Hang JF. Taipei Veterans General Hospital (VGHTE); Nat’l Yang Ming Chiao Tung University
STUDY DESIGN:
- 100 paired cytospin and CytoRich urine cytology slides from 50 bladder cancer patients digitized to create WSI and analyzed with AI-assisted software that identifies the most atypical urothelial cells and displays them in a gallery of the top 24
- 3 cytologists with variable digital pathology experience (A= over 12 months; B= 6-12 months; C= less than one month) reviewed the AI-assisted images and interpreted them using The Paris System categories
RESULTS:
- No significant difference in diagnostic performance of cytologists between preparations (cytospin vs CytoRich)
- Digital pathology experience improved performance, with C having the poorest performance
- Sensitivity = 84-96%; Specificity = 92- 96%; PPV = 91.3-95.7%; NPV = 85.2-95.8%; accuracy = 88-94%
- Cytologist C took the least time reviewing slides (median 29.5-30 sec) compared to the other 2 (median 63.5-71.5 sec)
CONCLUSIONS:
AI-assistance markedly improves efficiency for the interpretation of upper urinary tract cytology. Experience using the software system enhances overall performance and diagnostic concordance.
May 2023
AIxURO URINE CYTOLOGY CLINICAL STUDIES
iSMIT Conference Oral Presentation
A novel and effective deep-learning algorithm for risk stratification of urothelial carcinoma
A novel and effective deep-learning algorithm for risk stratification of urothelial carcinoma
International Society for Medical Innovation and Technology 2023 Oral Platform / Abstract (Min Inv Ther Allied Technol)
STUDY DESIGN:
- 116 urine cytology WSI (76 cytospin, 40 CytoRich) scanned with Leica AT2 at 20X and analyzed and ranked “top 24” by AI-deep learning algorithm for most abnormal urothelial cells
- 1 cytopathologist (CP) and 2 cytologists (CT) evaluated WSI with AI-assistance and diagnosed slides as NHGUC, AUC, SHGUC or HGUC according to The Paris System 2.0.
- Performance Metrics: Concordance with original cytologic diagnosis and time to diagnosis compared to conventional microscopy
RESULTS:
- AI-assistance increased sensitivity for all 2 reviewers (from 83.3-100%) and improved time efficiency from 159.9 mins to 106.3 min)
- Specificity was reduced for the CTs, but not the CP
CONCLUSIONS:
AI-assistance using a deep learning algorithm to identify and rank abnormal cells improves overall clinical sensitivity for the detection of urothelial carcinoma while improving clinical efficiency through time-savings.
March 2023
WHOLE SLIDE IMAGE SCANNER STUDIES
USCAP Poster Presentation
Optimization of the number of Z-stack layers for whole-slide images of urine cytology with different cellularity by an artificial intelligence algorithm
Optimization of the number of Z-stack layers for whole-slide images of urine cytology with different cellularity by an artificial intelligence algorithm
ASC/IAC 2022 Poster Presentation / Abstract Hang JF, Ou YC, Tsao TY et al. VGHTE, Tungs Taichung MetroHarbor Hospital
STUDY DESIGN:
- 14 positive (HGUC) urine cytology specimens (cytospin and ThinPrep) digitized with Leica Aperio AT2 scanner at 4 different scan modes:
- Mode A- Default automatic
- Mode B- Optimized automatic
- Mode C- Manually focused points of adjustment
- Mode D- Multilayer Z-stacking
- Performance metrics:
- Rate of successful scanning (total slides that obtain WSI files from scanner);
- Number of high risk cells detected by AI algorithm
- Coverage rate of high risk cells (detected high risk cells in one Z-layer divided by the total high risk cells detected in 21 Z-layer WSI per slide)
RESULTS:
- Advanced scan modes increase successful slide scanning, but does not increase the average high risk cell detection or coverage rates
- More high risk cells were detected with TP than cytospin but with comparable coverage rates
- More Z-layers increased average high risk cell detection and coverage rates
CONCLUSIONS:
Advanced scan modes improve the success rate of obtaining an optimal WSI and the number and coverage rate of high risk cells increases with multiple Z-layers. More high risk cells were detected on ThinPrep vs cytospin but this did not significantly affect the AI algorithm analytic results.
March 2023
SOFTWARE OPTIMIZATION & VALIDATION STUDIES
USCAP Poster Presentation
Optimization of snapshot gallery presentation of artificial intelligence-inferred diagnostic cell images in urine cytology
Optimization of snapshot gallery presentation of artificial intelligence-inferred diagnostic cell images in urine cytology
USCAP 2023 Poster Presentation / Abstract Liu TJ, Yang WL, Tung MC, Hang JF et al. Tungs Taichung MetroHarbor Hospital; VGHTE
STUDY DESIGN:
- 45 HGUC urine cytology slides and 45 negative (NHGUC) slides digitized into WSI and inferred by an AI-algorithm to identify potential cancer cells.
- 12, 24, 36, or 48 abnormal cell thumbnail images were order-ranked by AI and presented in a reviewer gallery
- 2 cytologists reviewed thumbnail images and rendered a diagnosis based on The Paris System for Reporting Urine Cytology 2.0 for each gallery set
- Concordance with microscopy performed for each gallery set, requiring an exact match (HGUC to HGUC, NHGUC to NHGUC)
RESULTS:
- Concordance for HGUC was significantly improved by increasing the number of diagnostic cells in the gallery from 38.9% to 91.1%, but were slightly decreased for NHGUC specimens from 92.2% to 90%.
- The highest concordance rate (95.6%) was established at 36 gallery cell images, but was only slightly decreased with 48 images
CONCLUSIONS:
36 thumbnail cell images provides the highest concordance for HGUC vs NHGUC diagnoses in urine cytology
November 2022
AIxURO URINE CYTOLOGY CLINICAL STUDIES
ASC/IAC Oral Presentation
Validation of a deep-learning-based algorithm for risk stratification and diagnosis of bladder cancer in digital urine cytology
Validation of a deep-learning-based algorithm for risk stratification and diagnosis of bladder cancer in digital urine cytology
ASC/IAC 2022 Oral Presentation / Abstract PLAT 15 Yang WL, Ou YC, Lin YS, Hang JF, Lui TJ. VGHTE, Taichung MetroHarbor Hospital
STUDY DESIGN:
- 104 urine cytology slides (70 NHGUC and 34 positive [AUC, SHGUC, HGUC]), with diagnostic confirmation by an expert panel
- Review of AI (deep learning algorithm)- inferred digital gallery images by 1 cytopathologist (CP) and 2 cytologists (CT), assigning cases to The Paris System (TPS 2.0) diagnostic categories (4 week washout period between Arm 1 and 3)
- Arm 1- Microscopic review of urine cytology glass slide by CP
- Arm 2- WSI review of slide by CP
- Arm 3- AI-assisted review of images by CP and CTs
- Performance Metrics:
- Comparison of diagnostic concordance with sensitivity, specificity, PPV, NPV and accuracy for CP and CTs against initial diagnosis
- Comparison of diagnostic concordance with sensitivity, specificity, PPV, NPV and accuracy between conventional microscopy and AI-assisted image review for the CP
RESULTS:
Microscopic Glass Slide Review versus AI-Assisted Image Review (CP only)
- Sensitivity: 85.2% vs 91.2% (+5.9%)
- Specificity: 100% vs 100%
- PPV: 100% vs 100%
- NPV: 93. vs 95.9% (+2.6%)
- Accuracy: 95.2% vs 97.1% (+1.9%)
AI-Assisted Review (CTA, CTB and CP)
- Sensitivity: 76.5%, 79.4%, 91.2%
- Specificity: 100%, 98.6%, 100%
- PPV: 100%, 96.4%, 100%
- NPV: 89.7%, 90.8%, 95.9%
- Accuracy: 92.3%, 92.3%, 97.1%
CONCLUSIONS:
AI-assisted diagnoses were comparable to expert panel consensus (sensitivity 76.5 – 91.2%) with high specificity (98.6-100%). AI-assisted interpretation showed better sensitivity, NPV, and accuracy than conventional microscopy.
November 2022
WHOLE SLIDE IMAGE SCANNER STUDIES
ASC/IAC Poster Presentation
Optimization of digital urine cytology via focus and Z-layer scan settings of the whole-slide imager
Optimization of digital urine cytology via focus and Z-layer scan settings of the whole-slide imager
ASC/IAC 2022 Poster Presentation / Abstract Hang JF, Ou YC, Tsao TY et al. VGHTE, Tungs Taichung MetroHarbor Hospital
STUDY DESIGN:
- 14 positive (HGUC) urine cytology specimens (cytospin and ThinPrep) digitized with Leica Aperio AT2 scanner at 4 different scan modes:
- Mode A- Default automatic
- Mode B- Optimized automatic
- Mode C- Manually focused points of adjustment
- Mode D- Multilayer Z-stacking
- Performance metrics:
- Rate of successful scanning (total slides that obtain WSI files from scanner);
- Number of high risk cells detected by AI algorithm
- Coverage rate of high risk cells (detected high risk cells in one Z-layer divided by the total high risk cells detected in 21 Z-layer WSI per slide)
RESULTS:
- Advanced scan modes increase successful slide scanning, but does not increase the average high risk cell detection or coverage rates
- More high risk cells were detected with TP than cytospin but with comparable coverage rates
- More Z-layers increased average high risk cell detection and coverage rates
CONCLUSIONS:
Advanced scan modes improve the success rate of obtaining an optimal WSI and the number and coverage rate of high risk cells increases with multiple Z-layers. More high risk cells were detected on ThinPrep vs cytospin but this did not significantly affect the AI algorithm analytic results.
June 2022
AIxURO URINE CYTOLOGY CLINICAL STUDIES
Academic Article: Cancer Cytopathology
Evaluation of an artificial intelligence algorithm for assisting the Paris System in reporting urinary cytology: A pilot study
Evaluation of an artificial intelligence algorithm for assisting the Paris System in reporting urinary cytology: A pilot study
Ou YC, Tsoa TY, Chang MC, et al. Cancer Cytopathology 2022; 130:872-880. Tung University Clinical Study
STUDY DESIGN:
- 131 urine cytology slides digitized and analyzed with AI-assisted algorithm for detecting and ranking abnormal urothelial cells using criteria from The Paris System (TPS2.0); ground truth diagnosis established by an expert panel
- Two arm study with 1 cytopathologist (CP) in both arms and 2 cytologists (CT) in the AI-assisted Arm 2
- Arm 1: CP review of urine cytology glass slide with diagnosis (4 week washout period)
- Arm 2: CP and CT review of AI-assisted inferred WSI of urine cytology slide with quantitative data statistics
- Performance Metrics: Comparison of research TPS diagnosis with expert diagnosis and calculation of sensitivity, specificity, PPV, NPV and accuracy
RESULTS:
- Conventional Microscopy vs AI-Assisted Interpretation
- Sensitivity: 87% vs 92.3% (+5.1)
- Specificity: 100% vs 100%
- PPV: 100% vs 100%
- NPV: 94.8% vs 96.8% (+2.0)
- AI-Assisted Microscopy, CTA vs CTB vs CP interpretation
- Sensitivity: 79.55 VS 82.1% VS 92.3%
- Specificity: 100% vs 98.9% vs 100%
- PPV: 100% vs 97% vs 100%
- NPV: 92% vs 92.9% vs 96.8%
- ĸ (95% CI): 0.845 vs 0.847 vs 0.994
CONCLUSIONS:
Ai-assisted interpretation was comparable to the expert consensus diagnosis, with superior sensitivity and NPV. AI-assisted interpretations showed near perfect agreement with the expert consensus diagnosis (ĸ = 0.944) and the microscopic diagnosis (ĸ = 0.862).
May 2021
AI SOFTWARE DEVELOPMENT
AACR Poster Presentation
Monitoring of stem cell differentiation to mature hepatocytes with a machine learning-based AI model
Monitoring of stem cell differentiation to mature hepatocytes with a machine learning-based AI model
AACR 2021 Poster Presentation / Abstract 185 Cancer Res (2021) 81 (13_Supplement): 185. https://doi.org/10.1158/1538-7445.AM2021-185 Yang WL, Huo Z, Chen SY, Zhu D, Liu TJ, Lee DF.
STUDY DESIGN:
- Develop a machine-learning-based artificial intelligence (AI) model to assist monitoring morphologic changes in human embryonic stem cells (hESC) in color, using bright field microscopy images
- Pilot Study: Train the model to estimate degree of stem cell differentiation at the Hepatic Progenitor Cell (HPC stage), the critical checkpoint for hepatocyte differentiation, based on cellular morphologic features
- Initial training set: Expert annotated images of 341 successful HPC differentiations and 366 failed HPC differentiations
- Cross-validation set: Images of 86 successful and 51 failed HPC results
- Test set: Images of 64 successful and 29 failed HPC results
- Failed differentiation = no differentiation or differentiation into non-hepatocyte tissue types
- Performance Metrics: Accuracy and F1 scores of test set
RESULTS:
The AI model showed excellent performance compared with the conventional method of determining degree of hepatocyte differentiation
- Accuracy = 0.978
- F1 score = 0.975
CONCLUSIONS:
AI-assisted models have the potential to improve the detection of degrees of hepatocyte differentiation, thereby improving the efficiency of a manual process that is very time-intensive.
Upcoming Events
Upcoming event
.
-
Barcelona, Spain
-
,
Latest news
November 2024
American Society of Cytopathology (ASC)
Dr. Barbara Crothers is the recipient of the ASC’s highest award- 2024 Papanicolaou AwardApril 2024
