MOSAIC: Explainable AI for Reproducible Histologic Grading and Prognostic Stratification in Breast Cancer

The MOSAIC framework is an explainable AI system that decomposes breast cancer histologic grading into its constituent components to significantly reduce inter-observer variability and improve prognostic stratification compared to traditional manual assessment.

The Gist

The Big Picture: The "Grading" Problem

Imagine a breast cancer diagnosis is like a report card for a tumor. Doctors need to give the tumor a "grade" (Grade 1, 2, or 3) to decide how aggressive it is and what treatment the patient needs.

The current system, called the Nottingham Grading System, is like a teacher grading a student's essay based on three specific things:

1. Tubule Formation: How organized is the tumor? (Is it neat like a library, or messy like a pile of laundry?)
2. Nuclear Pleomorphism: How weird do the cell nuclei look? (Are they all the same size, or are they bizarre and misshapen?)
3. Mitotic Activity: How fast are the cells dividing? (Are they sleeping, jogging, or sprinting?)

The Problem: This grading is done by human pathologists looking at microscope slides. It's very subjective. One doctor might see a cell dividing and say "Grade 2," while another looks at the same slide and says "Grade 3." This is like two teachers grading the same essay and giving it different scores because they are tired, the lighting is bad, or they just have different opinions. This inconsistency can lead to patients getting too much treatment or not enough.

The Solution: Enter MOSAIC

The researchers built MOSAIC (Mammary Oncology Spatial Analysis and Intelligent Classification). Think of MOSAIC not as a robot that replaces the doctor, but as a super-powered, tireless teaching assistant that sits next to the pathologist.

Instead of looking at the whole messy slide and guessing, MOSAIC breaks the problem down into three distinct tasks, just like the human grading system, but it does it with mathematical precision.

1. The "Tubule" Detective (The Organizer)

• Human way: The doctor squints at the slide, trying to find little ring-shaped structures (tubules) and counting them. It's hard to see them all.
• MOSAIC way: Imagine a drone flying over a city. MOSAIC scans the whole slide and instantly highlights every single "ring" structure, counts them, and calculates exactly how many are in a specific area. It doesn't get tired, and it doesn't miss a single one.

2. The "Mitosis" Counter (The Speedometer)

• Human way: The doctor has to find the cells that are in the middle of dividing. These look like tiny, dark dots. It's easy to miss them or mistake a shadow for a dividing cell.
• MOSAIC way: MOSAIC acts like a high-speed camera. It scans the slide, finds every single cell that is "sprinting" (dividing), and counts them. It even ignores the "fake runners" (cells that look like they are running but aren't). It then picks the 10 busiest areas to give a fair score.

3. The "Nucleus" Inspector (The Shape Shifter)

• Human way: The doctor looks at the size and shape of the cell centers (nuclei). Are they big? Are they weird? This is the hardest part to agree on because "weird" is subjective.
• MOSAIC way: MOSAIC measures the exact size of every single nucleus with a digital ruler. It calculates the average size and how much they vary. It's like a strict judge measuring the height of every contestant to the millimeter, removing all guesswork.

The "Teacher's Meeting" (The Study)

To test if this assistant actually helps, the researchers gathered 7 expert pathologists (the "teachers") and gave them a stack of 30 difficult slides to grade.

• Round 1 (No AI): They graded the slides alone. As expected, they disagreed with each other a lot. It was like a committee of teachers arguing over a grade.
• Round 2 (With AI): They graded the same slides again, but this time MOSAIC was sitting next to them, highlighting the dividing cells and measuring the shapes.
• The Result: The agreement between the doctors skyrocketed. The "teaching assistant" helped them see the same things. The doctors spent less time squinting and worrying, and their final grades were much more consistent.

The "Report Card" Results

The researchers also checked if MOSAIC's grades were better at predicting who would get sick again (prognosis).

• The Old Way: When doctors graded the slides manually, the "Grade 2" and "Grade 3" groups often looked the same on survival charts. It was hard to tell who was high risk.
• The MOSAIC Way: When MOSAIC did the grading, the groups separated clearly. The "high risk" group looked very different from the "low risk" group.
• The Metaphor: Imagine sorting a pile of rocks by size. A human might sort them into "Small," "Medium," and "Large," but the "Medium" pile is a mess of big and small rocks. MOSAIC sorts them so perfectly that the "Medium" pile is actually just the right size, making it much easier to predict which pile will cause a landslide (cancer recurrence).

Why This Matters

1. Fairness: A patient in Mumbai gets the same grade as a patient in New York, regardless of which doctor looks at their slide.
2. Speed: It helps doctors work faster, reducing burnout.
3. Clarity: It turns a "feeling" (this looks bad) into a "fact" (this has 14 dividing cells per area).

The Catch

The system isn't perfect yet. The "Nucleus Shape" part is still tricky for both humans and AI because it's inherently vague. Also, this is a preprint (a draft), so it needs more testing before it becomes standard in every hospital.

In short: MOSAIC is a smart, objective assistant that helps doctors grade breast cancer more consistently, ensuring patients get the right treatment based on clear, measurable facts rather than a tired doctor's guess.

Technical Summary

1. Problem Statement

Breast cancer prognosis and treatment planning rely heavily on the Nottingham Histologic Grade (NHG), which assesses three morphological features: mitotic activity, nuclear pleomorphism, and tubule formation. Despite its clinical importance, NHG assessment suffers from:

• High Inter- and Intra-observer Variability: Particularly in counting mitotic figures and assessing nuclear atypia, leading to inconsistent grading.
• Subjectivity: Manual grading is labor-intensive and prone to search bias and fatigue.
• Limited Prognostic Resolution: Traditional manual grading often fails to clearly stratify risk, especially between intermediate grades (Grade 1 vs. 2, or 2 vs. 3).
• Lack of Explainability: Many existing AI grading models use "end-to-end" deep learning that predicts a final grade without decomposing the decision into the specific histologic components required by pathologists, limiting trust and clinical adoption.

2. Methodology

The authors developed MOSAIC (Mammary Oncology Spatial Analysis and Intelligent Classification), an Explainable AI (xAI) framework designed to replicate the component-wise logic of human pathologists.

A. Architecture and Models

MOSAIC utilizes two complementary deep learning architectures applied to Whole Slide Images (WSIs):

1. Semantic Segmentation (U-Net):
   • Task: Delineates tissue compartments (Cellular Tumor, Stroma, Adipose, Necrosis, etc.) and identifies glandular structures.
   • Output: Used to quantify Tubule Formation by calculating tubule density specifically within cellular tumor regions.
2. Object Detection (DINO-DETR):
   • Task: Detects and classifies individual nuclei and mitotic figures.
   • Cell-Type Classification: Classifies nuclei into tumor cells, lymphocytes, macrophages, fibroblasts, etc., validated against 10x Genomics Xenium spatial transcriptomics data to ensure biological accuracy.
   • Mitotic Detection: Identifies mitotic figures and distinguishes them from "mitotic mimics" (non-dividing cells with similar morphology). It selects the 10 consecutive High-Power Fields (HPFs) with the highest mitotic activity for standardized scoring.

B. Calibration and Scoring Strategy

A unique Multi-Phase Pathology Calibration Study was conducted to bridge AI outputs with clinical standards:

• Phase 1 (Controlled): 7 expert pathologists graded 30 WSIs in four sessions (Baseline, Repeat, AI-Assisted, Repeat AI-Assisted). This established baseline variability and defined scoring thresholds.
• Threshold Derivation: Continuous AI-derived metrics were mapped to discrete NHG scores (1, 2, 3) based on consensus discussions and distribution analysis:
  • Mitotic: Based on count per 10 HPFs (or per 10,000 tumor cells if tumor area is small).
  • Nuclear Pleomorphism: Based on mean nuclear area and Interquartile Range (IQR).
  • Tubule Formation: Based on tubules per mm².
• Phase 2 (Large-Scale): Validated on 251 WSIs from the NIH-PLCO cohort to test scalability and reproducibility.

C. Datasets

• Training: TCGA-BRCA, MIDOG, TUPAC, NuCLS, and private Indian cohorts (SJRI, UL, PCCM).
• Validation: Public datasets (TCGA-BRCA, NIH-PLCO) and multiple independent Indian institutional cohorts (SJRI, Samved, Zydus, BVDUMC, UL).
• Biological Validation: Cross-validated using spatial transcriptomics (Xenium) and single-sample gene set enrichment analysis (ssGSEA).

3. Key Contributions

• Component-Wise Explainability: Unlike black-box models, MOSAIC outputs independent scores for mitotic activity, nuclear pleomorphism, and tubule formation, aligning directly with the Nottingham grading paradigm.
• Pathologist-in-the-Loop Calibration: The framework was not just trained on data but calibrated through a rigorous, multi-phase study with expert pathologists to define clinically relevant thresholds.
• Spatial Transcriptomics Integration: Used 10x Genomics Xenium data to validate that the AI's morphological cell-type classifications (e.g., tumor vs. immune cells) align with underlying gene expression profiles.
• Multi-Institutional Robustness: Validated across diverse datasets with varying staining protocols, scanners, and geographic locations (US and India) without dataset-specific recalibration.

4. Key Results

A. Technical Performance

• Tissue Segmentation: High accuracy (0.90) in distinguishing tumor from stroma.
• Tubule Formation: Highest agreement with pathologists (Accuracy: 0.66, $\kappa$ = 0.549).
• Mitotic Activity: Moderate agreement (Accuracy: 0.50, $\kappa$ = 0.40), but showed the most significant improvement in reproducibility when pathologists used AI assistance.
• Nuclear Pleomorphism: Lowest agreement (Accuracy: 0.33, $\kappa$ = 0.27), reflecting the inherent subjectivity of this feature even for humans.

B. Impact on Pathologist Consistency (Observer Study)

• Reduced Variability: AI assistance significantly improved inter- and intra-observer agreement, particularly for mitotic activity (weighted $\kappa$ increased to 0.98 in some cases) and tubule formation.
• Efficiency: AI assistance reduced average grading time per slide (e.g., from ~9 mins to ~4 mins for some pathologists) by reducing cognitive load and search bias.
• Stability: AI assistance stabilized grading decisions across repeated readings, especially for pathologists with lower baseline consistency.

C. Prognostic Stratification

• Superior Risk Separation: In the TCGA and NIH-PLCO cohorts, AI-derived grades provided significantly better survival stratification than manual pathologist grades.
  • TCGA: AI-derived grade yielded a global p-value of $5.9 \times 10^{-7}$ vs. non-significant results for manual grades in some endpoints.
  • NIH-PLCO: AI grading achieved a log-rank p < 0.0001, clearly separating Grade 1 from Grades 2 and 3, whereas manual grading showed overlap between Grade 1 and 2.
• Independent Prognostic Value: In multivariable Cox regression models (adjusting for age, stage, and subtype), AI-derived grade remained a significant independent predictor of survival, whereas manual NHG often lost significance.
• Hazard Ratios: AI-derived grades showed Hazard Ratios (HR) $\ge$ 1.8 in multiple datasets with $p < 0.001$.

5. Significance and Clinical Implications

• Reproducibility: MOSAIC addresses the "reproducibility crisis" in breast cancer grading by standardizing the assessment of the most variable components (mitosis and tubules).
• Decision Support, Not Replacement: The framework is designed as an assistive tool that preserves the pathologist's role while providing quantitative, visual evidence to guide difficult decisions.
• Cost-Effective Precision: By extracting high-dimensional prognostic features from standard H&E slides, MOSAIC offers a potential, lower-cost alternative to expensive genomic assays (like Oncotype DX) while retaining spatial context.
• Global Applicability: The successful validation in diverse Indian cohorts demonstrates the framework's robustness against variations in staining and scanning infrastructure, suggesting viability for low-resource settings.

Conclusion: MOSAIC represents a significant step forward in computational pathology by combining deep learning with rigorous clinical calibration. It transforms histologic grading from a subjective art into a reproducible, explainable, and prognostically superior quantitative process, ultimately supporting more consistent and effective breast cancer treatment planning.