AI-Based Pipeline for Vascularization Detection in Histological Images; low-cost solution Using Moderate Computing Resources

Technical Summary: AI-Based Pipeline for Vascularization Detection in Histological Images

Problem Statement
The manual assessment of vascularization in histological specimens is a critical indicator for wound healing and tissue response but remains a time-consuming process requiring significant expert effort. While artificial intelligence (AI) offers a potential solution, practical adoption is often hindered by high costs associated with licensing, the need for specialized AI personnel, and the requirement for powerful computing infrastructure (e.g., dedicated GPU servers). Furthermore, traditional deep learning approaches often demand large, fully annotated training datasets, which are costly to produce. The authors address the need for a low-cost, user-friendly, and computationally efficient system that can detect vascular structures in immunohistochemically stained tissue with minimal manual labeling and modest hardware requirements.

Methodology
The proposed pipeline is a retrieval-augmented, human-in-the-loop system designed to function on standard workstations without specialized GPU hardware. It consists of four logical stages:

1. Image Preprocessing and Candidate Mask Generation:

   • The system employs classical image processing to isolate smooth muscle actin (SMA) positive structures. This involves color channel separation (isolating the blue channel to suppress hematoxylin and emphasize DAB staining) and morphological filtering.
   • Two distinct methods generate candidate regions of interest (ROIs):
     • Ridge Detection: Implemented via the Fiji/ImageJ plugin (based on Steger's algorithm), this method enhances thin, elongated, curvilinear structures typical of vessel walls.
     • Segment Anything Model (SAM2): A zero-shot, training-free segmentation model (Meta AI) is used to generate diverse region proposals without task-specific retraining.
   • These methods produce binary masks or candidate regions that serve as inputs for feature extraction.

2. Region Embedding:

   • Each candidate ROI is cropped, normalized, and passed through a frozen, self-supervised visual backbone: DINOv2 (specifically dinov2_vits14).
   • DINOv2, pretrained on natural images, acts as a fixed descriptor extractor, converting each ROI into a compact numerical vector (embedding) that captures semantic and morphological characteristics. No further supervised fine-tuning of the backbone is performed.

3. Indexing and Retrieval-Based Classification:

   • The generated embeddings are stored in a dynamic vector index using FAISS (Facebook AI Similarity Search).
   • Classification is performed via a "retrieval-augmented" few-shot strategy. When a new ROI is presented, the system retrieves the top-k nearest neighbors from the index based on vector similarity.
   • The labels of these nearest neighbors (Vessel vs. Artifact) are used to vote on the classification of the new ROI.

4. Interactive Human-in-the-Loop Protocol:

   • The system operates in real-time. An expert reviewer views detected objects and provides immediate feedback via keyboard commands:
     • 'v' or 'a' keys: Explicitly label an object as a vessel or artifact. These labels are immediately added to the FAISS index, updating the system's knowledge base for subsequent images without retraining the underlying model.
     • 'Enter' or 's' keys: Used for evaluation (agreeing or disagreeing with the model's prediction) without updating the training set.
   • This allows the system to adapt incrementally to specific staining patterns and edge cases during a single session.

Key Contributions

• Low-Cost, Low-Resource Deployment: The authors demonstrate that high-accuracy vascular detection can be achieved on a standard desktop computer or gaming notebook, eliminating the need for expensive cloud GPU instances or dedicated servers required by traditional deep learning workflows.
• Annotation Efficiency: The retrieval-augmented approach achieves high accuracy (approx. 95%) with only ~75 user-assisted recognitions and stabilizes near 98% by 150 annotations. This is significantly fewer than the hundreds or thousands of annotations typically required for supervised neural network training.
• Real-Time Adaptability: Unlike static models that require retraining to incorporate new data, this system learns continuously and instantly from expert corrections, allowing rapid adaptation to specific laboratory conditions or staining variations.
• Complementary Segmentation: The study validates that combining classical ridge detection (which preserves continuous vascular morphology) with SAM2 proposals (which offer diverse shape coverage) provides a more robust set of candidates than either method alone.

Results
The study compared three approaches using histological images from a porcine wound healing experiment (SMA-stained):

1. Fully Manual Annotation: Required over 14,000 objects to be classified by a single operator over several months (2020–2022).
2. Standard Neural Network (Azure-based): A machine-learning tool hosted on Microsoft Azure reached 92% accuracy after 400 annotations but required substantial cloud computing resources and cost approximately 7 EUR for the annotation sessions.
3. Retrieval-Augmented System (DINOv2/FAISS):
   • Achieved 95.4% accuracy after only 75 assisted recognitions.
   • Reached 98.2% accuracy by 125 recognitions and plateaued near 98% at 150 recognitions.
   • Performed effectively on a standard workstation without dedicated GPUs.
   • Training Strategy Findings: The study found that distributing training data across multiple images (e.g., 26 objects from 13 images) yielded higher and more stable accuracy than training on objects from a single image. Furthermore, including morphologically ambiguous cases in the training set improved performance compared to restricting training only to unambiguous examples.

Significance and Claims
The paper claims that this approach offers a practical trade-off for routine laboratory use, particularly in settings where expert time is limited and computational resources are constrained. By leveraging pre-trained self-supervised representations (DINOv2) and efficient similarity search (FAISS), the system bypasses the need for task-specific training and large datasets.

The authors emphasize that the system is designed to assist, not replace, the expert; it reduces the annotation burden from months to minutes while maintaining high agreement with the operator. However, the paper modestly notes limitations:

• The system classifies detected objects as "vessel" or "artifact" but does not quantify the absolute number of vessels in an image.
• The accuracy metric reflects agreement with a specific human rater; if the rater makes systematic errors, the model will replicate them.
• Broader validation across different stains, tissue types, and laboratories is required before the method can be considered fully generalizable.

The work positions itself as a step toward label-efficient learning in medical imaging, demonstrating that a combination of self-supervised embeddings and interactive retrieval can effectively address the annotation constraints of routine pathology workflows.