Research Summary - 2
Supervised machine learning to diagnose respiratory disease-affected bovine lung histopathology photomicrographs
| Date/Time: |
8/28/2026
17:00
|
| Author: |
Matthew Scott |
| Clinic: |
Texas A&M University |
| City, State, ZIP: |
Canyon, TX 79015 |
H.M. Prosser, BS, MS
1
;
A. Finley, DVM, DACVP
2
;
B.J. White, DVM, MS
3
;
E.M. Bortoluzzi, DVM, MS, PhD
3
;
M.A. Scott, DVM, PhD
1
;
1Veterinary Education, Research, and Outreach Program, Texas A&M University, Canyon, TX, 79015
2Shreiber School of Veterinary Medicine, Rowan University, Mullica Hill, NJ, 0806
3Beef Cattle Institute, Kansas State University, Manhattan, KS, 66506
Introduction:
Respiratory disease in feedlot cattle includes bronchopneumonia (BP), interstitial pneumonia (IP), and bronchopneumonia with an interstitial pneumonia (BIP). These conditions, especially late in the feeding period (>45 days on feed), cause economic loss and welfare concerns. While gross necropsy diagnosis is helpful, histologic diagnosis is the gold standard to confirm presence of IP. Clinical and gross diagnoses are rarely confirmed by histology, which can mischaracterize the true frequency of each syndrome. This study evaluated the use of supervised machine learning (ML) to classify histologic lung images by disease type.
Materials and methods:
Lung samples (1cm2) from four regions were collected from 294 deceased feedlot cattle, fixed in formalin, processed, and evaluated by a board-certified veterinary anatomic pathologist. Samples were characterized as BP, IP, apparently healthy (NORMAL), and any condition that did not meet criteria for a BP, IP, or NORMAL diagnosis (OTHER). Digital slides were cropped into 4,168 x 4,168-pixel patches (~10X magnification) and labeled using the Microsoft Azure ML Studio (Microsoft Corporation, Redmond, WA, USA). Azure’s Automated ML studio image classification feature, with a ViTb16r22 vision transformer network, was used for modeling data (80% for training and 20% for testing).
Results:
Utilizing 6,329 patches (220 normal, 2,000 BP, 2,093 AIP, and 2,016 other), the model achieved 83.2% accuracy, with good sensitivity (83.2%) and specificity (94.4%); the model area under the curve (AUC) was 0.95. Across diagnoses, the sensitivity ranged from 80-97% and specificity was 90-99%.
Significance:
With further refinement, this approach could yield faster disease management and improved animal welfare in the feedlot setting.
