Optimizing Tissue Section Thickness for AI-Based Cell Classification

Introduction

Deep learning-based cell segmentation and classification seem destined to play a role in future histology. While recent advances have enabled accurate identification of individual cells in histological images, achieving pathologist-level performance remains challenging due to the variability between histological tissue sections.

Project Purpose

The project investigates a slide-by-slide human-in-the-loop approach, in which manually annotated cells from each new tissue section are temporarily added as training data before classifying the remaining cells. To identify the optimal conditions for accurate and efficient cell classification, six tissue section groups (Fig. 1), including five microsection thicknesses (2–10 μm) and tissue sections produced using an automated microtome (DS), were evaluated.

Results and Impact

The results demonstrated that tissue sections produced using an automated microtome (DS), followed by 2 μm tissue sections, achieved the highest classification accuracy while requiring fewer manually annotated training cells (Fig. 2). These findings support the development of more accurate and efficient AI-assisted digital pathology workflows.

Contact Information

Name: Ida Skovgaard Christiansen

Email: ida.skovgaard.christiansen@regionh.dk

Location: Department of Pathology, 5411

Position: Pathologist

Publications

Technical note: Impact of tissue section thickness on accuracy of cell classification with a deep learning network

Figure 1: Representative image tiles from the six tissue section groups evaluated in the study.

Figure 2: Classification performance of the deep learning model across the six tissue section groups. Tissue sections produced with an automated microtome (DS) achieved the highest classification accuracy, followed by 2 μm tissue sections.