AI-assisted diagnosis of animal diseases from medical images
A three-branch clinic spends up to two hours on a single thoracic radiograph description — the radiologist is overloaded, the MRI queue keeps growing. We deployed an AI assistant for preliminary reading. The system is trained on thousands of annotated veterinary images and can detect cardiomegaly, pleural effusion, lung patterns on thoracic X-rays, as well as artifacts on abdominal ultrasound.
We use EfficientNet-B4, ResNet-50, and DenseNet-121 architectures with fine-tuning on specialized datasets. Results are comparable to expert assessment by veterinary radiologists. Our experience in computer vision spans over 15 projects in medical and veterinary diagnostics. Deploying such a system typically pays off within 4–6 months by reducing consultation time and lowering the number of repeat visits. Savings on outsourcing radiologists can reach 50,000 RUB/month per clinic.
Why AI diagnosis of animal diseases from images is challenging
Differences from human radiology: anatomical variability across species (cat, dog, rabbit, horse, exotic), size range (Chihuahua vs. German Shepherd — body mass ratio 1:60), projections (dorsoventral instead of anteroposterior). A model trained on human images cannot be directly transferred.
Main image types:
- Radiographs (thorax, abdomen, skeleton)
- Ultrasound images (abdomen, heart, thyroid)
- MRI/CT (spine, head) — in large clinics
How the AI system diagnoses animal diseases from images
The most demanded task is evaluation of canine and feline thorax: cardiomegaly, pleural effusion, pneumonia, tumors.
Cardiomegaly — assessed via Vertebral Heart Score (VHS): total heart size in vertebrae. A regression model predicts VHS directly. On a dataset of 1,200 thoracic radiographs (dogs): MAE = 0.31 vertebrae vs. normal 9.7±0.5 for most breeds. Comparable to inter-examiner variability among veterinarians (0.4 vertebrae).
Pleural effusion — visually obvious pathology (loss of diaphragmatic contour clarity), but missed when small volume. Binary classifier ResNet-50: AUC = 0.94 on test set.
Lung patterns — bronchial, interstitial, alveolar patterns require multi-label classification (patterns can co-occur). EfficientNet-B4 with multi-label head: micro F1 = 0.82 on 5 pattern classes. EfficientNet-B4 outperforms ResNet-50 by 5% F1 on this task.
| Pathology | Model (AUC/MAE) | Veterinarian (inter-expert variability) |
|---|---|---|
| Cardiomegaly (VHS) | MAE 0.31 vertebrae | 0.4 vertebrae |
| Pleural effusion | AUC 0.94 | ~0.92 |
| Task | Model | Metric | Inference time |
|---|---|---|---|
| Lung pattern classification | EfficientNet-B4 | micro F1 0.82 | 45 ms |
| Cardiomegaly detection | DenseNet-121 | AUC 0.88 | 30 ms |
More about the datasets
- 1,200 canine thoracic radiographs with VHS and pathology annotations - 800 feline abdominal ultrasound images with artifact annotations - 200 canine spinal MRIs (hernias, compressions)Why data is the main challenge in veterinary CV
The primary problem in veterinary computer vision is the lack of public datasets comparable to medical scale. Our strategy:
- Partner with 2–5 clinics (500–2,000 labeled cases in 3–6 months)
- Transfer learning from ImageNet or medical datasets (ChestX-ray14) + fine-tuning on veterinary data
- Synthetic data: augmentation with realistic pathological overlays on normal images
Transfer learning from human thoracic radiographs to canine works: a DenseNet-121 pretrained on ChestX-ray14 after fine-tuning on 800 canine radiographs yields AUC = 0.88 for cardiomegaly — better than training from scratch on the same volume (AUC = 0.79).
One of our partners — a veterinary clinic chain — reduced average image description time from 40 to 12 minutes after deployment. This allowed a 30% increase in patient flow without hiring additional radiologists.
What is included in the work
- Documentation: architecture description, dataset, metrics, and annotation guide
- Trained model with weights and configuration
- REST API for integration into existing PACS or veterinary information system
- Inference scripts and code examples (Python, PyTorch)
- 3-month warranty support
- Staff training on system use (2–4 hours)
Timeline
Base system for thoracic radiographs (dogs/cats): 10–14 weeks. Extended module with ultrasound and multiple modalities: 18–26 weeks. Cost is calculated individually based on data volume and required functionality.
Contact us to evaluate your project. We guarantee a prototype on your data within 10 weeks. Get a consultation on adapting the model to your tasks.