TL;DR
This paper introduces a fully automated AI pipeline for detecting and counting mitoses in breast cancer whole slide images, aiming to overcome the variability of manual counting. The AI-based mitotic count (MC) demonstrated comparable prognostic value to traditional light-microscopic MC for predicting patient survival, suggesting its potential for clinical integration.
What Problem Does This Paper Address?
Accurately assessing tumor proliferation is crucial for predicting breast cancer patient outcomes. Mitotic count (MC) is a key component of histopathological grading, but manual counting is prone to significant inter- and intraobserver variability. This inconsistency can hinder its clinical utility and impact treatment decisions. While AI has shown promise in detecting mitoses, its prognostic value compared to traditional methods on large, real-world datasets, and its integration into clinical workflows, remained to be fully validated.
Methods
The researchers developed a fully automated pipeline for breast cancer survival prediction using AI. The core of this pipeline is a deep learning model, originally developed for the Mitosis Domain Generalization Challenge 2021 (MIDOG21), which detects mitotic figures (MFs) in whole slide images (WSIs). To address the requirement of counting mitoses within a specific "hotspot" area (2 mm²), they introduced a novel automatic area selector. This selector uses two methods: an overlapping window search and a "bounded area maximum enclosing convex hull" (BAME convex hull) algorithm, designed to mimic how a pathologist identifies the most mitotically active region. The pipeline first detects all potential MFs, filters out those outside the tumor annotation, and then applies the area selector to find the optimal 2 mm² region to calculate the final MC.
Datasets
The study utilized a large dataset of 912 breast cancer patients from the University Medical Centre Utrecht with long-term follow-up data. This cohort was curated from a larger group of 2,230 patients, with specific inclusion criteria such as WSI scanned at ×40 magnification, availability of survival data, and acceptable image quality. The AI model itself was initially developed using a dataset from the MIDOG21 challenge.
Key Insights and Ideas
- AI as a Reproducibility Enhancer: The primary motivation is to leverage AI to standardize the MC assessment, mitigating the subjective nature of manual counting and improving reproducibility.
- Prognostic Equivalence: A key finding is that the AI-generated MC is not just comparable in detection but also in its ability to predict patient survival, matching the prognostic power of traditional MC.
- Automated Hotspot Identification: The development of an automatic area selector is crucial for clinical translation, as it removes the need for manual identification of the mitotic hotspot, a time-consuming and subjective step.
- Clinical Workflow Integration: The paper highlights the importance of integrating AI tools into existing clinical systems (like PACS) for practical adoption and evaluation.
Key Results
- The AI-based MC showed a strong correlation with the light-microscopic MC from the original pathology reports (Pearson r = 0.58, p < 0.00001).
- Both the AI-based BAME convex hull and the window search methods demonstrated comparable prognostic value to the traditional light-microscopic MC in uni- and multivariate survival analyses.
- Kaplan-Meier survival curves for the three MC score classes (derived from both traditional and AI methods) showed significant divergence, indicating prognostic stratification.
- In a small subgroup of cases reviewed by a pathologist, AI output correction led to an improved correlation with the original report MC.
Limitations and Future Work
- Hotspot Focus: The AI model was primarily trained on hotspot regions and not entire WSIs, which might make it less robust to artifacts or detections outside the tumor area. The authors suggest that integrating a tumor segmentation model could help focus the AI on relevant regions.
- Pathologist Interaction: While the AI performs well, the authors emphasize that pathologist interaction and review of AI output are crucial for clinical practice. They propose presenting AI results in a way that facilitates easy review and correction.
- Further Validation: The authors suggest that future studies should address multi-observer reproducibility and the economic impact of time savings for laboratories.
- Data Availability: The specific patient data used in this study is not publicly available due to identifying information.
Summary updated