Deep learning-based automated Nottingham histological grading of canine invasive mammary carcinoma with a preliminary cross-species exploration on feline mammary carcinoma

The Nottingham histological grading system (NHGS) for canine invasive mammary carcinoma relies on visual assessment, and inter-observer agreement remains below the accepted threshold for reliability, being lowest for nuclear pleomorphism (NP). We developed a deep learning system that scores all three NHGS components (tubule formation, TF; mitotic count, MC; and NP) on whole-slide images (WSIs) of confirmed canine invasive mammary carcinoma, with each score derived from a quantitative measurement rather than subjective assessment. A transformer-based segmentation network first delineates the mammary hyperplasia region on each WSI and produces a binary mask, which serves as a common spatial constraint for all three scoring modules. For TF, the glandular ratio (GR), defined as the proportion of tubule-structure pixels within the masked region, is computed at pixel level and replaces the pathologist's visual estimate of tubule proportion. For MC, mitotic figures are detected on image patches cropped from the masked region, and a greedy algorithm then selects the 10 densest non-overlapping high-power fields, thereby avoiding subjective selection of the counting area. For NP, a binary classifier detects patches containing nuclei fulfilling the score-3 morphological criteria, and the slide-level NP score is assigned from the proportion of positive patches, with the decision threshold calibrated on the validation set. The system was evaluated on 29 independent canine WSIs against the consensus of two senior veterinary pathologists. Seven feline mammary carcinoma WSIs, none of which were used in training and validation, were included as a preliminary cross-species exploration. Overall grading accuracy was 93.1% and the Gwet's AC1 was 0.919. Per-component agreement was 82.8% (TF), 86.2% (MC), and 72.4% (NP); given the markedly skewed score distributions, Gwet's AC1 was also calculated (0.771, 0.838, and 0.649). On the feline slides, overall accuracy reached 85.7%, with a bias pattern consistent with that observed in canine slides. This study was a feasibility investigation. The system generated NHGS grades without manual intervention, and the feline tissue results suggest that this method may be transferable to other companion animal species.

Authors

Institutions

Publication Details

Journal
Frontiers in Veterinary Science
Published
2026-11-19
DOI
https://doi.org/10.3389/fvets.2026.1998039
Primary Topic
AI in cancer detection
Type
article
Field-Weighted Citation Impact
0.00
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
article

Deep learning-based automated Nottingham histological grading of canine invasive mammary carcinoma with a preliminary cross-species exploration on feline mammary carcinoma

Yukun Guo, Hewei Zhang, Tong Wei, Jinfeng Zhu et al.
Frontiers in Veterinary Science
AI in cancer detection
article

Deep learning-based automated Nottingham histological grading of canine invasive mammary carcinoma with a preliminary cross-species exploration on feline mammary carcinoma

Yukun Guo, Hewei Zhang, Tong Wei, Jinfeng Zhu, Jie Wang, Yongjie Feng, Baowen Li
article en

Abstract

The Nottingham histological grading system (NHGS) for canine invasive mammary carcinoma relies on visual assessment, and inter-observer agreement remains below the accepted threshold for reliability, being lowest for nuclear pleomorphism (NP). We developed a deep learning system that scores all three NHGS components (tubule formation, TF; mitotic count, MC; and NP) on whole-slide images (WSIs) of confirmed canine invasive mammary carcinoma, with each score derived from a quantitative measurement rather than subjective assessment. A transformer-based segmentation network first delineates the mammary hyperplasia region on each WSI and produces a binary mask, which serves as a common spatial constraint for all three scoring modules. For TF, the glandular ratio (GR), defined as the proportion of tubule-structure pixels within the masked region, is computed at pixel level and replaces the pathologist's visual estimate of tubule proportion. For MC, mitotic figures are detected on image patches cropped from the masked region, and a greedy algorithm then selects the 10 densest non-overlapping high-power fields, thereby avoiding subjective selection of the counting area. For NP, a binary classifier detects patches containing nuclei fulfilling the score-3 morphological criteria, and the slide-level NP score is assigned from the proportion of positive patches, with the decision threshold calibrated on the validation set. The system was evaluated on 29 independent canine WSIs against the consensus of two senior veterinary pathologists. Seven feline mammary carcinoma WSIs, none of which were used in training and validation, were included as a preliminary cross-species exploration. Overall grading accuracy was 93.1% and the Gwet's AC1 was 0.919. Per-component agreement was 82.8% (TF), 86.2% (MC), and 72.4% (NP); given the markedly skewed score distributions, Gwet's AC1 was also calculated (0.771, 0.838, and 0.649). On the feline slides, overall accuracy reached 85.7%, with a bias pattern consistent with that observed in canine slides. This study was a feasibility investigation. The system generated NHGS grades without manual intervention, and the feline tissue results suggest that this method may be transferable to other companion animal species.

Frontiers in Veterinary ScienceVol. 13
Jiangsu Vocational College of Agriculture and Forestry (CN), Luoyang Polytechnic (CN)
AI in cancer detection
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon