Z-Stack Scanning can Improve AI Detection of Mitosis: A Case Study of Meningiomas

Hongyan Gu, Ellie Onstott, Wenzhong Yan, Tengyou Xu, Ruolin Wang, Zida Wu, Xiang Anthony Chen, Mohammad Haeri

Abstract

Identifying mitotic figures is a critical component of tumor grading in pathology, yet it remains a challenging task due to the limitations of single-layer digital imaging. This study investigates the efficacy of "z-stack" scanning—capturing multiple focal planes—to enhance artificial intelligence (AI) performance in detecting mitoses within meningioma tumors. By evaluating a two-stage deep learning pipeline across three different high-end scanners, the researchers demonstrate that z-stacking significantly boosts sensitivity by an average of 17.14% with virtually no loss in precision. While the approach increases storage requirements by approximately 3.81×, the results suggest that 3D spatial information is vital for overcoming the "focus gap" in digital pathology.

Introduction

In clinical pathology, the mitotic count is a cornerstone of tumor grading and patient management, particularly for meningiomas, the most common primary brain tumor. Detecting these tiny (~10µm) cellular structures is notoriously difficult; human pathologists often exhibit low concordance, and traditional AI models are frequently hampered by the quality of digital scans.

Most digital pathology workflows rely on "single-layer" whole slide imaging (WSI), which captures only one focal plane. If a mitotic figure is slightly out of focus or possesses complex 3D chromosomal structures, it may be missed or misidentified as a mimic. While "z-stack" scanning (capturing multiple planes along the z-axis) has been proposed as a solution, its clinical value has remained a subject of debate. This research seeks to provide a systematic evaluation of z-stacking across various hardware and software architectures to determine its true utility in automated mitosis detection.

Methodology

The researchers developed a robust, two-stage deep learning pipeline designed to compare the performance of single-layer versus z-stack imaging directly.

  1. Candidate Selection (Segmentation): In the first stage, a segmentation model (tested using PSPNet, Segformer, and DeepLabV3+) scans the WSI to identify potential mitotic candidates. For z-stack data, the model processes every layer and merges candidates that appear across multiple planes to ensure spatial consistency.
  2. Verification (Ensemble): The second stage utilizes an ensemble of four Convolutional Neural Networks (EfficientNet-b3, b5, v2-s, and v2-m) to analyze the candidates. In the z-stack workflow, this ensemble generates 20 distinct predictions (5 planes × 4 models) for each candidate. These predictions are then aggregated by a Random Forest classifier to produce a final probability score.
Comparison of the single-layer (a) and z-stack (b) mitosis detection pipelines.

To ensure the findings were hardware-agnostic, the pipeline was validated using three different high-end scanners: the Pannoramic 480DX, Aperio GT 450, and AxioScan 7.

Datasets and Calibration

The evaluation was conducted on 22 H&E-stained meningioma slides from the University of Kansas Medical Center, containing 6,350 expert-annotated mitoses. The models were pre-trained on a massive multi-source dataset of 38,634 mitoses, including data from MIDOG++ and MITOS WSI CMC. To account for imaging variations between different scanner vendors, one slide per scanner was reserved to re-calibrate the Random Forest classifier.

Key Results and Insights

The study's primary finding is that z-stacking bridges the "focus gap" that often plagues digital pathology. By allowing the AI to "twiddle the focus knob" digitally, the model can perceive nuanced chromosomal features that are blurred in a standard 2D slice.

  • Sensitivity Gains: Across all nine combinations of scanners and models, sensitivity increased from an average of 0.601 to 0.704, representing a 17.14% improvement. The highest sensitivity achieved was 0.773 using the DeepLabV3+ model on GT 450 scans.
  • Precision Stability: Unlike many sensitivity-boosting techniques that increase false positives, z-stacking provided a "free lunch." Average precision remained stable (moving from 0.753 to 0.757), indicating that the extra spatial data helped the AI confirm real mitoses rather than hallucinating new ones.
  • Visual Clarity: The multi-plane data proved particularly useful for identifying mitoses that were out of focus in the primary scan layer, as shown in visual comparisons across different hardware platforms.
Visual comparison of mitotic figures across single-layer and multiple z-planes using three different scanners.

Discussion and Limitations

While the performance gains are substantial, they come with a significant computational and storage cost. Z-stack files in this study were approximately 3.81× larger than their single-layer counterparts (418.92 GB vs. 87.02 GB). This ballooning of data presents a hurdle for widespread clinical adoption and real-time processing.

Future research should investigate "extended focus" or "all-in-focus" algorithms that might collapse multiple z-layers into a single high-detail 2D image. Such technologies could potentially retain the AI performance gains of z-stacking while maintaining manageable file sizes. Additionally, while this study focused on meningiomas, the benefits of z-stacking should be validated in other "low-burden" pathologies, such as the detection of H. pylori or the analysis of cytopathology smears.

Summary updated

BibTeX
@inproceedings{gu2025z,
  author = {Hongyan Gu and Ellie Onstott and Wenzhong Yan and Tengyou Xu and Ruolin Wang and Zida Wu and Xiang Anthony Chen and Mohammad Haeri},
  title = {{Z-Stack Scanning can Improve AI Detection of Mitosis: A Case Study of Meningiomas}},
  booktitle = {ISBI},
  year = {2025},
  doi = {10.1109/isbi60581.2025.10980734},
}
Figure