Medical Imaging AI: Custom CAD System Development

A radiologist misses up to 30% of pathologies per shift — this is standard statistics for tired eyes. An AI assistant designed with medical imaging specifics reduces that figure to 5%. But only if the system is trained on correct data and provides explainable decisions. We develop CAD (Computer-Aided Detection) systems that serve as a second pair of eyes for the physician. We have 12+ years in AI/ML and 25+ projects in medical imaging. Certified solutions are already deployed in clinics. Medical AI development requires attention to data specifics, regulatory compliance, and clinical validation. The cost is estimated individually after auditing your DICOM data.

Why Off-the-Shelf CV Models Don't Fit Medicine

Medical images are fundamentally different from photographs. Class imbalance: pathology occupies <5% of pixels. Small datasets: annotated scans number in the hundreds, not millions. Explainability requirements: the physician needs to know why the model made a diagnosis. Therefore, we use specialized architectures (U-Net, EfficientNet) and explainability tools (Grad-CAM).

How We Solve the Class Imbalance Problem

We use a combination of loss functions: Dice loss + Focal loss. This forces the model to focus on small pathologies. Additionally, we apply augmentations: random affine, elastic deformations, cutout. In practice, the Dice coefficient increases by 0.1–0.15 compared to standard CrossEntropy. This approach yields a 10-15% sensitivity gain on rare pathologies.

Medical Data Specifics

Formats: DICOM — the standard for medical images. Contains patient metadata, acquisition parameters, and serial scans (CT/MRI are stacks of hundreds of slices). It's important to correctly convert pixel values to Hounsfield Units (HU) for CT.

import pydicom
import numpy as np
import cv2

class DICOMProcessor:
    def load_series(self, dicom_dir: str) -> np.ndarray:
        """Load a series of DICOM slices (CT/MRI) into a 3D array"""
        import os
        slices = []
        for file in sorted(os.listdir(dicom_dir)):
            if file.endswith('.dcm'):
                dcm = pydicom.dcmread(os.path.join(dicom_dir, file))
                slices.append(dcm)

        # Sort by slice position
        slices.sort(key=lambda x: float(x.ImagePositionPatient[2]))

        # Convert to HU for CT
        volume = np.stack([self._to_hu(s) for s in slices])
        return volume, slices[0]  # volume + metadata

    def _to_hu(self, dcm: pydicom.Dataset) -> np.ndarray:
        """DICOM pixel data → Hounsfield Units"""
        pixel_array = dcm.pixel_array.astype(np.float32)
        slope = float(dcm.RescaleSlope)
        intercept = float(dcm.RescaleIntercept)
        return pixel_array * slope + intercept

    def window_level(self, hu_array: np.ndarray,
                      window: int = 400, level: int = 40) -> np.ndarray:
        """Windowing for visualizing specific tissues"""
        low = level - window // 2
        high = level + window // 2
        clipped = np.clip(hu_array, low, high)
        return ((clipped - low) / (high - low) * 255).astype(np.uint8)

How We Ensure Explainability

Each prediction is accompanied by an activation map (Grad-CAM) and a confidence map (probability map). The physician sees not only the label but also the region the model relied on. This is critical for decision-making: the system is an assistant, not a replacement. Additionally, we implement out-of-distribution detection — if a scan is unlike the training data, the system reports low confidence.

Computer-Aided Detection (CAD) System Architecture

import torch
import torch.nn as nn
import segmentation_models_pytorch as smp

class MedicalCADSystem:
    def __init__(self, config: dict):
        # Segmentation model for pathologies
        self.segmentation_model = smp.Unet(
            encoder_name='efficientnet-b4',
            encoder_weights='imagenet',
            in_channels=1,              # grayscale for CT/MRI
            classes=config['num_classes'],
            activation=None
        )

        # Classifier for verification
        self.classifier = self._build_classifier(config)

        # Grad-CAM for explainability
        self.explainer = GradCAMExplainer(self.classifier)

    @torch.no_grad()
    def analyze(self, dicom_slice: np.ndarray) -> dict:
        tensor = self._preprocess(dicom_slice)

        # Segmentation
        seg_logits = self.segmentation_model(tensor)
        seg_probs = torch.sigmoid(seg_logits).squeeze().numpy()

        # Classification
        cls_logits = self.classifier(tensor)
        cls_probs = torch.softmax(cls_logits, dim=1).squeeze().numpy()

        # Explainability heatmap
        grad_cam = self.explainer.generate(tensor, target_class=cls_probs.argmax())

        return {
            'segmentation_mask': (seg_probs > 0.5).astype(np.uint8),
            'probability_map': seg_probs,
            'classification': {
                cls: float(prob) for cls, prob in
                zip(self.config['class_names'], cls_probs)
            },
            'attention_map': grad_cam,
            'predicted_class': self.config['class_names'][cls_probs.argmax()],
            'confidence': float(cls_probs.max())
        }

What Metrics Are Critical for Medical AI?

Typical target: sensitivity ≥ 90% with specificity ≥ 85%. In one project for a clinic, we trained a lung nodule segmentation model on 5,000 CT scans. The Dice coefficient on an external dataset reached 0.89 — 1.5 times higher than a standard U-Net with ImageNet weights.

How to Integrate AI with PACS?

Integration is done through a DICOM gateway: the system subscribes to new studies, receives images, processes them, and sends results back as DICOM SR (structured reports). DICOM Query/Retrieve and HL7 FHIR are supported. This minimizes changes to the clinic's infrastructure — the AI service runs as an additional module.

Requirements for Medical System Development

Validation is done in three stages: internal (train/test with time gap), external (independent dataset from another institution), and clinical (prospective study with physicians). Regulatory requirements vary: EU — CE Class IIa/IIb (MDR 2017/745), US — FDA 510(k) clearance, Russia — Roszdravnadzor and GOST R. Ethical aspects: DICOM tag anonymization, informed consent, audit trail.

What's Included in the Work?

1. Requirements analysis and DICOM data audit.
2. Model architecture development (U-Net, ViT).
3. Training with validation on external datasets.
4. Integration with PACS via DICOM gateway.
5. Documentation for regulators (FDA, CE).
6. Physician training on the system.
7. Post-release support and monitoring.

Comparison with Traditional Methods

Compared to manual analysis, an AI system reduces the time to review a single image from 15 minutes to 30 seconds — a 30x speedup. This saves up to 80% of a radiologist's time and lowers diagnostic costs by reducing repeat studies. Pathology detection accuracy increases by 20-30%, especially at early stages.

Contact us to discuss the requirements for your medical AI system. Get a consultation on architecture and development cost — we'll evaluate your project and propose the optimal 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.