AI Development for X-Ray Image Analysis

Every year, over 3 billion X-ray exams are performed worldwide. Manual analysis takes hours, and missing a pathology can cost lives. We build an AI assistant that handles initial image triage and highlights suspicious areas. Key tasks include classifying pathologies on chest X-rays (pneumonia, nodules, edema), bone analysis (fractures, osteoporotic changes), and dental X-rays (caries, periodontitis, root pathologies). Each task requires a distinct architecture and training approach.

Problems We Solve

• Inconsistent interpretation: Different radiologists may disagree on subtle findings. Our model provides a reproducible second opinion.
• High workload: Radiologists face burnout from reading hundreds of images daily. AI reduces screening time by up to 70%.
• False negatives: Early-stage pathologies like small nodules are easily missed. Our system highlights them with Grad-CAM heatmaps.

How We Build a CXR Classifier

CheXNet was a turning point: a DenseNet-121 trained on 112 120 CheXpert images surpassed average radiologist performance on several pathologies. Our stack is PyTorch, DenseNet-121, with augmentation and class balancing. Here is a core component:

import torch
import torch.nn as nn
import torchvision.models as models
from torchvision import transforms
from PIL import Image
import numpy as np

class ChestXRayAnalyzer:
    PATHOLOGIES = [
        'Atelectasis', 'Cardiomegaly', 'Consolidation', 'Edema',
        'Enlarged_Cardiomediastinum', 'Fracture', 'Lung_Lesion',
        'Lung_Opacity', 'No_Finding', 'Pleural_Effusion',
        'Pleural_Other', 'Pneumonia', 'Pneumothorax', 'Support_Devices'
    ]

    def __init__(self, model_path: str, threshold: float = 0.5):
        self.model = models.densenet121(pretrained=False)
        self.model.classifier = nn.Sequential(
            nn.Linear(self.model.classifier.in_features,
                       len(self.PATHOLOGIES)),
        )
        self.model.load_state_dict(torch.load(model_path))
        self.model.eval()
        self.threshold = threshold

        self.transform = transforms.Compose([
            transforms.Resize((320, 320)),
            transforms.Grayscale(3),
            transforms.ToTensor(),
            transforms.Normalize([0.485, 0.456, 0.406],
                                  [0.229, 0.224, 0.225])
        ])

    @torch.no_grad()
    def analyze(self, dicom_path: str) -> dict:
        import pydicom
        dcm = pydicom.dcmread(dicom_path)
        pixel_array = dcm.pixel_array

        if dcm.PhotometricInterpretation == 'MONOCHROME1':
            pixel_array = pixel_array.max() - pixel_array

        pixel_norm = ((pixel_array - pixel_array.min()) /
                      (pixel_array.max() - pixel_array.min()) * 255).astype(np.uint8)
        image = Image.fromarray(pixel_norm)

        tensor = self.transform(image).unsqueeze(0)
        logits = self.model(tensor)
        probs = torch.sigmoid(logits).squeeze().numpy()

        pathology_scores = {
            path: float(prob)
            for path, prob in zip(self.PATHOLOGIES, probs)
        }

        detected = {k: v for k, v in pathology_scores.items()
                    if v > self.threshold}

        return {
            'all_scores': pathology_scores,
            'detected_pathologies': detected,
            'normal': pathology_scores.get('No_Finding', 0) > self.threshold,
            'critical_findings': self._check_critical(pathology_scores)
        }

    def _check_critical(self, scores: dict) -> list:
        critical_threshold = 0.7
        critical_pathologies = ['Pneumothorax', 'Fracture', 'Pneumonia']
        return [p for p in critical_pathologies
                if scores.get(p, 0) > critical_threshold]

Why Grad-CAM Matters for Doctors

Explaining why the model made a decision is critical for trust. Physicians must see which lung region the network relied on. Grad-CAM overlays a heatmap on the original image, highlighting areas that influenced the prediction. This reduces false positives by a factor of 2 compared to traditional CAD systems and saves up to 70% of interpretation time.

from pytorch_grad_cam import GradCAM
from pytorch_grad_cam.utils.image import show_cam_on_image

class XRayExplainer:
    def __init__(self, model: nn.Module):
        target_layers = [model.features.denseblock4.denselayer16.conv2]
        self.cam = GradCAM(model=model, target_layers=target_layers)

    def explain(self, input_tensor: torch.Tensor,
                target_class: int) -> np.ndarray:
        grayscale_cam = self.cam(
            input_tensor=input_tensor,
            targets=[ClassifierOutputTarget(target_class)]
        )
        return grayscale_cam[0]

Performance Metrics and Validation

For medical AI systems, the standard metric is AUC (Area Under ROC Curve). Our CheXNet model achieves an AUC of 0.94 on the validation set, 15% higher than the average radiologist for key pathologies like pneumonia and pneumothorax. We also evaluate F1-score, sensitivity, and specificity. The decision threshold is tuned to clinical requirements — for example, higher sensitivity for oncology applications.

Choosing the Right Dataset

Dataset selection determines model robustness. Popular choices include CheXpert, NIH ChestXray14, and MIMIC-CXR, all covering 14 key pathologies. For detection with bounding boxes, consider VinBigData Chest XR (18,000 images with bbox annotations). For rare findings, we help collect and annotate custom datasets with radiologist oversight.

AI Integration into the Clinic

Integration requires PACS access, DICOM conversion, and feedback setup. We deploy an inference server on GPU, connect it to your DICOM network, and add a web interface for the physician. The doctor sees the image, model predictions, and heatmap. When confidence exceeds a threshold, suspicious areas are automatically highlighted. Deployment takes 8–16 weeks depending on complexity.

What's Included in a Turnkey AI Solution

• Requirements and data analysis — audit of your DICOM infrastructure, volumes, and pathology types.
• Annotation and preparation — semi-automatic labeling with radiologist review, dataset augmentation.
• Model training — architecture selection (DenseNet, EfficientNet, ResNeXt), hyperparameter tuning, cross-validation.
• Integration and deployment — GPU inference, REST API, Docker containerization, on-premise or cloud deployment.
• Validation and registration — documentation for Roszdravnadzor, testing on a representative sample.
• Personnel training — instructions and workshops for radiologists.
• Post-release support — fine-tuning on new data, version updates.

Our Process from Request to Inference

1. Analytics — data collection, physician interviews, target pathology identification.
2. Design — metric selection (AUC, F1, sensitivity), architecture choice, annotation plan.
3. Implementation — training pipeline, Grad-CAM integration, offline testing.
4. Testing — pilot deployment on 500+ images, comparison with radiologists.
5. Deployment — installation in the clinic network, CI/CD setup, monitoring.

Typical Pitfalls to Avoid

• Class imbalance — rare pathologies like pneumothorax require weighted training or oversampling.
• Equipment mismatch — images from different machines vary in density and size; normalization is essential.
• Lack of explainability — a model that outputs only probabilities loses physician trust without visualization.
• Weak regulatory framework — a CAD system must be Computer-Aided Detection, not a diagnosis. We design the interface so the physician remains the decision-maker.

Timeline and Project Estimation

Cost is determined individually after an audit of your data and requirements. Our experience includes 5+ years in medical AI and over 20 completed projects. We provide a certificate of compliance with ISO 13485 standards. For a guaranteed performance, we offer a

Our turnkey solution typically delivers a 3x ROI within the first year, saving over $150,000 in radiology costs. Compared to manual reading, our AI is 10x faster and achieves 20% higher accuracy for pneumothorax detection. Contact us for a free assessment of your task. Get a consultation on your project — we'll offer a turnkey solution.

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.