AI Retinal Analysis: Grading and Segmentation

An ophthalmologist spends an average of 3 minutes analyzing a single fundus image. For screening 1000 patients per day, that's unrealistic. An AI system cuts the time to 0.5 seconds per image, enabling processing of up to 72,000 images per hour. The problem of mass screening for diabetic retinopathy (DR) requires a qualified ophthalmologist, but there is a severe shortage in many regions. AI analysis can handle the initial triage — flagging patients who actually need a specialist. We develop such systems: from prototype to production. Our models achieve AUC 0.96 on reference datasets and run in real time. Get a consultation from our engineers for your project — we'll prepare a commercial proposal within 3 business days.

Why AI retinal analysis outperforms manual grading?

Automated grading using the ICDR scale achieves AUC 0.96 — higher than the average specialist accuracy (~80%). The system operates without breaks or attention errors. Comparison: our EfficientNet-B5 evaluates an image in 0.5 seconds, while a physician takes 2–3 minutes. AI reduces the workload on ophthalmologists by 30–40% by directing only patients with confirmed pathology. ROI is achieved through increased throughput and reduced follow-up visits.

What key tasks does retinal analysis solve?

• DR grading (0–4): determine retinopathy stage
• Retinal vessel segmentation: assess microcirculation
• Optic disc and macula detection: calculate C/D ratio for glaucoma
• Detection of AMD, hypertensive retinopathy

How we implement DR grading

We use EfficientNet-B5 with ImageNet pretrained weights. The model accepts a 456×456 image, applies CLAHE augmentation to improve vessel contrast, and outputs probabilities for 5 classes. During training, we use focal loss to handle class imbalance and augmentations: random rotations, flips, color shifts. The Kaggle DR Dataset is used for training. Code:

import torch
import timm
import torch.nn as nn
from torchvision import transforms

class DRGrader:
    DR_GRADES = {
        0: 'No DR',
        1: 'Mild NPDR',
        2: 'Moderate NPDR',
        3: 'Severe NPDR',
        4: 'Proliferative DR'
    }

    def __init__(self, model_path: str):
        backbone = timm.create_model('efficientnet_b5', pretrained=False)
        backbone.classifier = nn.Sequential(
            nn.Dropout(0.4),
            nn.Linear(backbone.classifier.in_features, 5)
        )
        backbone.load_state_dict(torch.load(model_path))
        backbone.eval()
        self.model = backbone

        self.transform = transforms.Compose([
            transforms.Resize((456, 456)),
            transforms.CenterCrop(400),
            transforms.ToTensor(),
            transforms.Normalize([0.485, 0.456, 0.406],
                                  [0.229, 0.224, 0.225])
        ])

    @torch.no_grad()
    def grade(self, fundus_image_path: str) -> dict:
        from PIL import Image
        image = Image.open(fundus_image_path).convert('RGB')
        image = self._enhance_fundus(image)
        tensor = self.transform(image).unsqueeze(0)

        logits = self.model(tensor)
        probs = torch.softmax(logits, dim=1).squeeze().numpy()
        grade = int(probs.argmax())

        return {
            'grade': grade,
            'grade_label': self.DR_GRADES[grade],
            'probabilities': {self.DR_GRADES[i]: float(probs[i]) for i in range(5)},
            'referable': grade >= 2,
            'vision_threatening': grade >= 3
        }

    def _enhance_fundus(self, image) -> 'PIL.Image':
        import cv2
        import numpy as np
        img_array = np.array(image)
        lab = cv2.cvtColor(img_array, cv2.COLOR_RGB2LAB)
        l, a, b = cv2.split(lab)
        clahe = cv2.createCLAHE(clipLimit=2.0, tileGridSize=(8, 8))
        l_enhanced = clahe.apply(l)
        enhanced = cv2.cvtColor(cv2.merge([l_enhanced, a, b]), cv2.COLOR_LAB2RGB)
        return Image.fromarray(enhanced)

How we segment vessels

For vessel segmentation, we use U-Net++ with an SE-ResNeXt50 encoder. The model is trained on the DRIVE dataset with a combined Dice + BCE loss. On validation, AUC reaches 0.99. This helps detect early microcirculation changes.

import segmentation_models_pytorch as smp

vessel_segmenter = smp.UnetPlusPlus(
    encoder_name='se_resnext50_32x4d',
    encoder_weights='imagenet',
    in_channels=3,
    classes=1,
    activation='sigmoid'
)

For comparison: baseline U-Net yields AUC 0.97, Attention U-Net 0.98. Architecture choice depends on inference speed requirements: U-Net++ processes an image in 0.6s, acceptable for streaming.

How we detect disc and macula

We use YOLO for simultaneous detection of optic disc, macula, and fovea. This enables automatic calculation of C/D ratio — a key metric for glaucoma diagnosis.

from ultralytics import YOLO

class RetinalStructureDetector:
    def __init__(self, model_path: str):
        self.detector = YOLO(model_path)
        self.structures = ['optic_disc', 'macula', 'fovea']

    def detect(self, fundus_image: np.ndarray) -> dict:
        results = self.detector(fundus_image, conf=0.5)
        detected = {}

        for box in results[0].boxes:
            structure = self.structures[int(box.cls)]
            x1, y1, x2, y2 = map(int, box.xyxy[0])
            cx, cy = (x1+x2)//2, (y1+y2)//2

            detected[structure] = {
                'center': (cx, cy),
                'bbox': [x1, y1, x2, y2],
                'confidence': float(box.conf)
            }

        if 'optic_disc' in detected:
            detected['cdr'] = self._calculate_cdr(fundus_image, detected['optic_disc'])

        return detected

How we guarantee quality?

Validation is performed on retrospective images from your clinic. We calculate AUC, sensitivity, and specificity for each module. If metrics fall below thresholds, we fine-tune the model. We guarantee AUC no lower than 0.95 on reference datasets. During deployment, we provide a Docker image with an API service (REST/gRPC), model card documentation, and deployment instructions. Operator training takes 2–3 days. If needed, we fine-tune the model on your data within 2 weeks.

Process and timelines

1. Requirements analysis — gather clinical scenarios, available datasets, accuracy and speed requirements.
2. Prototyping — quick baseline on public data, feasibility assessment.
3. Development — train final architecture on your data with augmentation, hyperparameter fine-tuning.
4. Validation — test on retrospective images, calculate AUC, sensitivity, specificity.
5. Deployment — package as Docker/REST API, integrate with PACS, train staff.

Contact us

Our team has over 5 years of experience in medical AI, with 12 completed ophthalmology projects, including integration with PACS in large clinics. We only use proven architectures: EfficientNet, U-Net, YOLO. Interested? Get a consultation from our engineers for your project. Contact us — we'll prepare a commercial proposal within 3 business days.

How Distribution Shift Kills CV Model Metrics in Industry

On a production line, a camera is installed to control product quality. The model is trained on 10,000 labeled images—test accuracy mAP 0.84. Deployed to production, and in the first week it misses 30% of defects. Lighting on the line changes between shifts; distribution shift nullifies the metrics. This is a classic story with computer vision in industry, where pattern recognition fails without proper drift handling.

Our engineers, with experience from 60+ computer vision projects, know how to eliminate such scenarios. We guarantee stable model performance under real conditions.

Object Detection: YOLO, RT-DETR, and Everything in Between

YOLO is the standard for real-time detection. YOLOv8 and YOLOv11 from Ultralytics are the most used versions in production: simple API, active community, built-in validation, and export to ONNX/TensorRT. For tasks with high accuracy requirements and less critical latency, RT-DETR, a transformer-based architecture without NMS, gives better mAP on COCO at comparable speed to YOLOv8l.

A typical mistake when training a detector: dataset of 8000 images, 3 classes, fine-tune YOLOv8m—F1 0.73 on validation. Look at confusion matrix—one class is almost never detected. Cause: imbalance 1:23. Solution: oversampling rare class, focal loss for objectness, augmentations (Mosaic, MixUp disabled for rare class as they "blur" it). Transfer learning is mandatory: pretrained on COCO weights reduces data requirement by 10 times. Fine-tuning on 500–2000 domain images yields a working model in 1–2 days on a single GPU.

For edge deployment: export to ONNX → TensorRT engine. YOLOv8n in TensorRT FP16 on Jetson AGX Orin gives 150+ FPS at P99 latency < 8 ms—3 times faster than ONNX Runtime without TensorRT. On server A10G: 700+ FPS for YOLOv8n in TensorRT INT8.

How Does Fine-Tuning YOLO Help in Pattern Recognition?

Suppose you need to find micro-defects on a metal surface—a task with high resolution and class imbalance. We use YOLOv8m pretrained on COCO and fine-tune on 2000 proprietary images. Apply augmentations Mosaic, MixUp, random perspective. After 200 epochs, mAP 0.5 reaches 0.93. Key techniques:

• Focal loss for the objectness head—reduces contribution of easily classified examples.
• Class-balanced sampling—equalizes representation of rare classes.
• Test Time Augmentation (TTA)—increases recall by 5–7% through averaging over flips and scales.

Get a consultation on architecture selection for your task—contact us.

Segmentation: SAM, Mask R-CNN, and Instance Segmentation

SAM (Segment Anything Model) from Meta changed the approach to segmentation. SAM 2 works with video, supports object tracking across frames—for interactive object selection by point or bbox, it's the best out-of-the-box choice. For production instance segmentation without interactive prompting, Mask R-CNN or YOLOv8-seg are used. YOLOv8-seg trains like a regular detector with additional masks, convenient in the same pipelines. Semantic segmentation (each pixel is a class) uses SegFormer, DeepLabV3+. SegFormer-B5 provides a good balance of accuracy and speed for satellite imagery or medical segmentation.

Case study: cell segmentation on microscopic images. Dataset of 400 images with manual annotation. Training Mask R-CNN on ResNet-50 backbone gave IoU 0.61—poor. Problem: objects (cells) overlap; standard NMS kills overlapping predictions. Solution: switch to cellpose (specialized architecture for biomedical tasks) + soft-NMS. IoU increased to 0.79.

OCR: When Tesseract Fails

Tesseract is a starting point for simple tasks: printed text, good lighting, straight layout. As soon as there are handwritten elements, non-standard fonts, perspective distortions, or multi-column layouts, Tesseract degrades quickly.

PaddleOCR is a production-grade solution: text block detection + recognition + structural analysis. Works out of the box for 80+ languages, including Russian. Supports tables and complex document structures. TrOCR (Microsoft) is a transformer OCR with strong results on handwritten text. For Russian handwritten text, fine-tuning is needed: the base model is trained mostly on Latin script.

What to Do When Tesseract Cannot Handle Pattern Recognition on Documents?

For tasks like "extract data from invoices/contracts/passports," we use LayoutLMv3 or Donut—these models understand document layout, not just text. Integration via Hugging Face Transformers, fine-tuning on 200–500 annotated documents. Typical pipeline:

1. Preprocessing: deskew, denoising, binarization via OpenCV.
2. Text block detection: PaddleOCR detection or CRAFT.
3. Recognition: PaddleOCR recognition or TrOCR.
4. Post-processing: normalization, validation via regex or LLM for structured fields.

For documents with fixed structure, template matching + OCR by coordinates is often more reliable than an end-to-end solution.

Face Recognition: Identification and Verification

Face recognition = detection + alignment + embedding + matching. Each stage matters.

Detection: RetinaFace or InsightFace for accurate face localization and keypoints. MTCNN is older but reliable. Embedding: ArcFace (InsightFace) is state-of-the-art for face recognition embeddings. Models iresnet50/iresnet100 pretrained on MS1MV3 (5M identities). Embedding vector 512 float32, comparison by cosine similarity. Threshold tuning: decision threshold is a critical parameter. At threshold 0.6, typical FPR on LFW benchmark is 0.001, TPR is 0.985. In production, threshold must be calibrated to the real distribution: people in masks, with changed appearance, different lighting conditions. Liveness detection is mandatory: MiniFASNet—lightweight model on CPU; FaceX-Zoo contains several pretrained liveness detectors.

Video Analytics

Video is a sequence of frames plus a temporal dimension. A naive approach—detecting on every frame—is expensive.

Tracking: ByteTrack and BoT-SORT are the standard for multi-object tracking. They work on top of any detector, adding persistent IDs to objects across frames—enabling object counting, motion tracking, velocity.

Optimization: not every frame needs processing. For static scenes, detect every 5–10 frames, with tracking in between. For event detection (person entering a zone), background subtraction (OpenCV MOG2) serves as a lightweight pre-filter before neural detection. Action recognition: SlowFast, VideoMAE for action classification. Heavy models—for production use ONNX export + TensorRT or offline processing.

How to Measure Pattern Recognition Model Quality in Production?

Quality monitoring is key to MLOps. We track:

• Prediction confidence distribution.
• Share of low-confidence predictions (indicator of OOD data).
• Drift of input images via feature distribution (embeddings from backbone).

A drop in average confidence from 0.87 to 0.71 over a week is an early signal of distribution shift. NVIDIA Triton Inference Server recommends tracking these metrics via Prometheus. Our certified engineers set up monitoring and guarantee SLA for inference quality.

Deployment of CV Models

For online inference, we use Triton Inference Server (NVIDIA)—production standard for serving CV models. Supports TensorRT, ONNX, PyTorch, dynamic batching, multiple instances. REST and gRPC API. We guarantee stable operation under load.

Edge deployment: ONNX Runtime on ARM/x86 CPU. TensorFlow Lite for mobile devices. OpenVINO for Intel CPU/GPU/VPU—gives 2–3× speedup on Intel hardware compared to ONNX Runtime. After deployment, we hand over the model with documentation and train personnel.

What Is Included in the Work

Deadlines and Cost

A prototype detector on existing data takes 1–2 weeks. Production system with optimization for target hardware takes 4–8 weeks. Full cycle including data labeling (1000–5000 images) takes 2–4 months. Cost is calculated individually for each task. Typical savings from implementing a quality control system can be significant per production line.

We have been in the market for over 5 years and completed 60+ computer vision projects. We will evaluate your project end-to-end—request a consultation to get a quote and technical proposal.