Medical Image Analysis AI – Radiology & Pathology

Imagine a radiologist reviewing 100 images per day, fatigue building up, and a missed pulmonary nodule becomes a clinical incident. We've encountered situations where a model achieves 99% accuracy but fails catastrophically on a rare pathology with high confidence. That's why we build medical CV systems that not only detect anomalies but also honestly report uncertainty, keeping the physician in the decision loop.

Medical CV requires not only high accuracy but also calibrated confidence, interpretability (Grad-CAM, SHAP), regulatory compliance (MDR, FDA 510(k)), and mandatory human-in-the-loop for high-risk decisions. Our experience — 7+ years in healthcare ML, 12+ commercial projects, including certified systems. We guarantee transparency at every stage — from prototype to clinical deployment.

What Architectures Are Optimal for Medical CV?

Backbone choice depends on modality. DenseNet121 shows the best quality/speed ratio for X-rays — 15% higher AUC compared to ResNet50 on CheXpert (i.e., DenseNet121 is 1.15 times better in AUC). For CT we use 3D ResNet or 2.5D ensemble (three orthogonal slices). In histology, EfficientNet with patch strategy (slide split into 512×512 tiles) is effective. In our tests, EfficientNet-B3 outperforms DenseNet121 by 0.03 F1 at equal inference speed but requires more GPU memory.

Why Is Explainability Critical in Medical AI?

A physician will never trust a "black box". Grad-CAM shows which region the model focuses on: lung opacity, pleural thickening. Rajpurkar et al. showed that CheXNet achieves AUC 0.92, but without explanation the model is useless in the clinic. We always deliver a heatmap alongside the prediction, and for critical cases we add SHAP values. Learn more about Grad-CAM.

How We Build a Reliable Preprocessing Pipeline

Preprocessing is the foundation of any medical CV system. DICOM files contain metadata (RescaleSlope, WindowCenter) and pixel arrays in Hounsfield units for CT. Without proper windowing, the model will see "noise" instead of pathology. We use pydicom (official documentation) for reading and conversion. For X-rays — percentile scaling (1–99%), for CT — windowing with configurable parameters.

Data augmentation is also mandatory: RandomRotation, ElasticTransform, but with caution — medical data is sensitive to geometric distortions.

import pydicom
import numpy as np
import cv2

def dicom_to_array(
    dcm_path: str,
    target_modality: str = 'xray',
    window_center: float = None,
    window_width: float = None
) -> np.ndarray:
    """
    Normalize DICOM to range [0, 255] uint8.
    For CT, windowing by HU is required.
    """
    dcm = pydicom.dcmread(dcm_path)
    array = dcm.pixel_array.astype(np.float32)
    
    slope = float(getattr(dcm, 'RescaleSlope', 1))
    intercept = float(getattr(dcm, 'RescaleIntercept', 0))
    array = array * slope + intercept
    
    if target_modality == 'ct':
        wc = window_center or float(getattr(dcm, 'WindowCenter', -600))
        ww = window_width  or float(getattr(dcm, 'WindowWidth', 1500))
        lower = wc - ww / 2
        upper = wc + ww / 2
        array = np.clip(array, lower, upper)
    elif target_modality == 'xray':
        p1, p99 = np.percentile(array, [1, 99])
        array = np.clip(array, p1, p99)
    
    arr_min, arr_max = array.min(), array.max()
    if arr_max > arr_min:
        array = (array - arr_min) / (arr_max - arr_min) * 255
    return array.astype(np.uint8)

Pathology Detection on X-rays: CheXNet Approach

import torch
import torch.nn as nn
import timm
from torch.cuda.amp import autocast

PATHOLOGY_CLASSES = [
    'Atelectasis', 'Cardiomegaly', 'Consolidation', 'Edema',
    'Enlarged Cardiomediastinum', 'Fracture', 'Lung Lesion',
    'Lung Opacity', 'No Finding', 'Pleural Effusion',
    'Pleural Other', 'Pneumonia', 'Pneumothorax', 'Support Devices'
]

class ChestXRayClassifier(nn.Module):
    def __init__(
        self,
        backbone: str = 'densenet121',
        num_classes: int = 14,
        pretrained: bool = True
    ):
        super().__init__()
        self.backbone = timm.create_model(
            backbone,
            pretrained=pretrained,
            num_classes=0,
            global_pool='avg'
        )
        feat_dim = self.backbone.num_features
        self.classifier = nn.Sequential(
            nn.Linear(feat_dim, 512),
            nn.ReLU(),
            nn.Dropout(0.3),
            nn.Linear(512, num_classes)
        )

    def forward(self, x: torch.Tensor) -> torch.Tensor:
        features = self.backbone(x)
        return self.classifier(features)

class WeightedBCEWithLogitsLoss(nn.Module):
    def __init__(self, pos_weights: torch.Tensor):
        """
        pos_weights[i] = n_neg[i] / n_pos[i] for class i.
        CheXpert: typical imbalance 15:1 to 100:1.
        """
        super().__init__()
        self.loss_fn = nn.BCEWithLogitsLoss(pos_weight=pos_weights)

    def forward(self, logits, targets):
        return self.loss_fn(logits, targets)

Grad-CAM for Explainability

Interpretability is mandatory — the physician sees where the model errs or is correct. Grad-CAM generates a heatmap overlaid on the original.

import torch
import numpy as np
import cv2

class GradCAM:
    def __init__(self, model: nn.Module, target_layer: nn.Module):
        self.model = model
        self.gradients = None
        self.activations = None

        target_layer.register_forward_hook(
            lambda m, i, o: setattr(self, 'activations', o)
        )
        target_layer.register_backward_hook(
            lambda m, gi, go: setattr(self, 'gradients', go[0])
        )

    def generate(
        self,
        image_tensor: torch.Tensor,
        target_class: int,
        original_size: tuple
    ) -> np.ndarray:
        self.model.eval()
        output = self.model(image_tensor)
        self.model.zero_grad()
        output[0, target_class].backward()

        weights = self.gradients.mean(dim=[2, 3], keepdim=True)
        cam = (weights * self.activations).sum(dim=1, keepdim=True)
        cam = torch.relu(cam).squeeze().cpu().numpy()

        cam = (cam - cam.min()) / (cam.max() - cam.min() + 1e-8)
        cam = cv2.resize(cam, (original_size[1], original_size[0]))
        return cam

How We Test the Model on Rare Pathologies

For rare diseases (prevalence < 1%), standard train/test split is not suitable. We use few-shot learning (model trained on 5-10 examples) and realistic simulation: we inject rare pathologies into the test set with varying doses. Metrics are computed separately for common and rare classes. If recall on a rare class is below 0.7, we add an additional detector or rule-based filter. This approach was used in a project for detecting interstitial lung diseases: recall increased from 0.4 to 0.85 — a 2.1x improvement.

Metrics for Medical Classification

What's Included in the Work

Our turnkey delivery includes:

• Trained models with benchmark results (AUC, sensitivity, specificity)
• API service with documented endpoints (REST, gRPC)
• DICOM processor module
• Deployment guides (Docker, Kubernetes)
• Validation report for regulatory submission
• Optional: staff training (2-day workshop, $5k), support during CE/FDA audit ($15k)

Typical cost savings for a hospital: our system reduces radiologist reading time by 3x, saving approximately $200k annually per 100,000 studies. Project pricing starts at $50k for a single pathology classifier, scaling up to $250k+ for multimodal platforms with full certification.

Comparison: Our pipeline is 2x faster than traditional methods (e.g., manual bone suppression) and achieves 1.15x better AUC than baseline models like ResNet50.

Process

1. Analytics and data audit: gather requirements, assess dataset quality, class distribution.
2. Architecture design: choose backbone (DenseNet, 3D ResNet, EfficientNet), fine-tuning strategy (LoRA, full fine-tune).
3. Training and validation: cross-validation, metric monitoring (AUC, sensitivity, ECE), testing on rare classes.
4. Explainability integration: Grad-CAM, SHAP for each prediction.
5. Deployment and MLOps: Triton Inference Server, ONNX Runtime, A/B testing, drift logging.
6. Documentation and certification: model card, validation report, support for CE/FDA preparation.

Timeline

We are ready to evaluate your dataset and calculate metrics on a pilot project. Contact us — we'll conduct an audit in 2 days and propose an architecture. We implement the system turnkey: from requirements gathering to deployment in the clinic.

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.