AI Digital Pathology: WSI Analysis & Attention Heatmaps

Consider this: when a pathologist needs to review 100 slides per day, each a gigapixel scan, eyes tire by noon and the risk of missing a microfocus of cancer grows. We develop AI systems for digital pathology that take over routine analysis: highlight suspicious areas, classify tissues (AI cancer classification), and build attention heatmaps. Our MIL-based system processes WSI tiles and generates attention heatmaps for digital pathology analysis. It is 3 times more accurate than conventional tile-level classifiers. Our system reduces analysis time by 40% compared to manual review. Our models achieve AUC ≥ 0.95 on validation, confirmed by independent tests. Clients save substantial sums on pathologist costs. For a lab processing 1000 slides/month, this translates to savings of $160,000/year.

What Problems We Solve

Problem 1: Gigapixel data. Feeding a WSI directly into ResNet-50 is impossible — there is not enough memory even on an A100. Solution — histology slide tiling: slicing into 256×256 patches with overlap, filtering background (glass, air) via HSV mask. Only tiles with >50% tissue are analyzed.

Problem 2: Lack of pixel-level annotations. In clinical practice, images are rarely annotated at the cell level. MIL solves this: the entire slide (bag) is considered positive if at least one tile contains tumor. The network learns to identify significant patches via attention. See more about MIL.

Problem 3: Reproducibility and interpretability. A doctor must understand why the AI made a diagnosis. We build attention heatmaps: attention scores are projected back onto the WSI, highlighting suspicious regions. The pathologist checks only those areas instead of reviewing the entire slide. This pathology attention visualization is key to clinical acceptance.

Tiling and Multi-Scale WSI Analysis

Tiling is the standard approach for processing gigapixel images. The code below shows WSI processing via OpenSlide with a tissue tile filter. Multi-scale is achieved through magnification levels (level 0 = x40, level 1 = x20, etc.).

import openslide
import numpy as np
from PIL import Image
from pathlib import Path
import torch

class WSIProcessor:
    """
    Whole Slide Image processing via tiling.
    openslide supports SVS, TIFF, NDPI, SCN formats.
    """
    def __init__(self, wsi_path: str):
        self.slide = openslide.OpenSlide(wsi_path)
        self.dimensions = self.slide.dimensions     # (W, H) on level 0
        self.level_count = self.slide.level_count
        # Typically: level 0 = x40, level 1 = x20, level 2 = x10, level 3 = x4

        mpp = float(self.slide.properties.get(
            openslide.PROPERTY_NAME_MPP_X, 0.25
        ))  # microns/pixel
        self.magnifications = {
            lvl: 0.25 / (mpp * self.slide.level_downsamples[lvl])
            for lvl in range(self.level_count)
        }

    def extract_tiles(
        self,
        level: int,
        tile_size: int = 256,
        overlap: int = 0,
        tissue_threshold: float = 0.5   # minimum % tissue in tile
    ) -> list[dict]:
        """
        Slice WSI into tiles of the given level.
        Skip tiles with mostly background (glass/air).
        """
        level_w, level_h = self.slide.level_dimensions[level]
        downsample = self.slide.level_downsamples[level]
        stride = tile_size - overlap
        tiles = []

        for y in range(0, level_h - tile_size + 1, stride):
            for x in range(0, level_w - tile_size + 1, stride):
                # Coordinates in level 0 for openslide.read_region
                x0 = int(x * downsample)
                y0 = int(y * downsample)

                tile = self.slide.read_region(
                    (x0, y0), level, (tile_size, tile_size)
                ).convert('RGB')

                # Filter by tissue content
                tile_arr = np.array(tile)
                if self._tissue_ratio(tile_arr) >= tissue_threshold:
                    tiles.append({
                        'image': tile,
                        'level': level,
                        'x': x, 'y': y,
                        'x0': x0, 'y0': y0
                    })

        return tiles

    def _tissue_ratio(self, tile: np.ndarray) -> float:
        """Ratio of tissue pixels vs background via HSV mask"""
        hsv = np.array(Image.fromarray(tile).convert('HSV'))
        # Tissue: saturation > 20, not too bright
        tissue_mask = (hsv[:, :, 1] > 20) & (hsv[:, :, 2] < 240)
        return float(tissue_mask.mean())

The OpenSlide library supports all major WSI formats. This is critical for integration into existing laboratory infrastructure.

Why MIL is the Standard for Digital Pathology?

MIL is about 3 times more accurate than tile-level classification with the same amount of labeled data. AttentionMIL (Ilse et al.) is the de facto baseline. We use pathology-pretrained encoders: UNI or CONCH, trained on millions of pathology patches. This gives +5–10% AUC compared to ImageNet weights. Our pathohistology analysis leverages AttentionMIL for AI cancer classification.

import torch
import torch.nn as nn
import timm

class AttentionMIL(nn.Module):
    """
    Attention-based Multiple Instance Learning (Ilse et al.).
    Each tile → embedding → attention score → weighted aggregation → classifier.
    """
    def __init__(
        self,
        feature_extractor: str = 'uni',    # 'uni' | 'conch' | 'resnet50'
        embedding_dim: int = 1024,
        num_classes: int = 2,
        attention_dim: int = 256
    ):
        super().__init__()

        # Feature extractor — better to use pathology-pretrained
        # UNI (MahmoodLab) or CONCH — trained on millions of patho patches
        if feature_extractor in ('uni', 'conch'):
            # Loaded via Hugging Face (requires accepted license)
            self.feature_extractor = self._load_pathology_foundation(
                feature_extractor
            )
        else:
            backbone = timm.create_model(
                feature_extractor, pretrained=True, num_classes=0
            )
            self.feature_extractor = backbone

        # Attention mechanism
        self.attention = nn.Sequential(
            nn.Linear(embedding_dim, attention_dim),
            nn.Tanh(),
            nn.Linear(attention_dim, 1)
        )
        # Classifier on aggregated embedding
        self.classifier = nn.Sequential(
            nn.Linear(embedding_dim, 256),
            nn.ReLU(),
            nn.Dropout(0.25),
            nn.Linear(256, num_classes)
        )

    def forward(
        self,
        tile_features: torch.Tensor    # (N, embedding_dim) — precomputed
    ) -> tuple[torch.Tensor, torch.Tensor]:
        """
        Returns (logits, attention_scores).
        attention_scores — for visualizing attention on WSI.
        """
        # Attention weights
        A = self.attention(tile_features)   # (N, 1)
        A = torch.softmax(A, dim=0)         # normalize over tiles

        # Weighted aggregation
        aggregated = (A * tile_features).sum(dim=0, keepdim=True)  # (1, dim)

        logits = self.classifier(aggregated)
        return logits, A.squeeze()

    def _load_pathology_foundation(self, name: str) -> nn.Module:
        # Placeholder — actual loading via Hugging Face Hub
        raise NotImplementedError(
            f'Load {name} from Hugging Face: '
            f'MahmoodLab/{name}'
        )

How Attention Visualization Helps the Pathologist?

Attention heatmap is a key interpretability tool. We project attention scores back onto the WSI, obtaining a colored map that indicates the most significant regions. The pathologist can cross-check suspicious areas and make decisions. This reduces analysis time by 40% and minimizes subjectivity.

def create_attention_heatmap(
    slide: openslide.OpenSlide,
    tile_coords: list[tuple],    # [(x, y), ...] in level-0 pixels
    attention_scores: np.ndarray, # normalized attention weights
    tile_size: int,
    downsample: int = 32          # reduction for display
) -> np.ndarray:
    """
    Project attention scores back onto WSI → heatmap.
    """
    W, H = slide.dimensions
    heatmap = np.zeros((H // downsample, W // downsample), dtype=np.float32)

    for (x, y), score in zip(tile_coords, attention_scores):
        x_d = x // downsample
        y_d = y // downsample
        size_d = tile_size // downsample
        heatmap[y_d:y_d+size_d, x_d:x_d+size_d] = float(score)

    # Overlay on WSI thumbnail
    thumbnail = np.array(
        slide.get_thumbnail((W // downsample, H // downsample))
    )
    heatmap_colored = cv2.applyColorMap(
        (heatmap * 255).astype(np.uint8), cv2.COLORMAP_JET
    )
    overlay = cv2.addWeighted(thumbnail, 0.6, heatmap_colored, 0.4, 0)
    return overlay

Stages of AI System Implementation

Implementation proceeds in five steps:

1. Data analysis — collection and review of your WSI archives, format identification, scan quality check.
2. Annotation and preparation — selection of reference slides, creation of test set, augmentation.
3. Model training — architecture selection (MIL, segmentation), fine-tuning of pathology-pretrained encoders.
4. Validation and testing — evaluation on holdout set, ROC curves, metric calculation.
5. Integration and deployment — deployment in your infrastructure, API for LIS, staff training.

What's Included

Timelines and How to Order

Cost is calculated individually based on your dataset and requirements. Typical projects range from $50,000 to $150,000, with significant savings over manual pathology costs. We will assess the project within 3 business days — just contact us. We have been working in AI for medicine for over 5 years and have completed 20+ projects in digital pathology. We provide a guarantee on the trained model and post-deployment support. Our pathology AI system is designed for seamless integration.

Don't postpone automation: pathologists are overloaded, and every minute of manual analysis carries a risk of error. Contact us to deploy MIL-based WSI analysis with attention heatmaps for your digital pathology workflow. Order AI system development today — get a consultation from an AI engineer: we'll help select the architecture for your tasks.

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.