🤖 machine learning

INTERACT-CMIL: Multi-Task Shared Learning and Inter-Task Consistency for Conjunctival Melanocytic Intraepithelial Lesion Grading

The paper introduces INTERACT-CMIL, a multi-task deep learning framework that leverages shared feature learning and inter-task consistency to significantly improve the accurate grading of Conjunctival Melanocytic Intraepithelial Lesions across five histopathological axes, outperforming existing CNN and foundation model baselines on a multi-center dataset.

Original authors: Mert Ikinci, Luna Toma, Karin U. Loeffler, Leticia Ussem, Daniela Süsskind, Julia M. Weller, Yousef Yeganeh, Martina C. Herwig-Carl, Shadi Albarqouni

Published 2026-05-27

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Mert Ikinci, Luna Toma, Karin U. Loeffler, Leticia Ussem, Daniela Süsskind, Julia M. Weller, Yousef Yeganeh, Martina C. Herwig-Carl, Shadi Albarqouni

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

The Big Picture: Grading a "Mole" on the Eye

Imagine a doctor looking at a tiny speck of skin on the eye (a conjunctival lesion). They need to decide: Is this harmless, or is it a dangerous cancer?

In the past, doctors had to look at this under a microscope and make a judgment call based on five different "rules" or criteria. It's like trying to grade a student's essay not just on grammar, but also on creativity, structure, vocabulary, and originality—all at the same time. The problem is that these rules are tricky. Sometimes one rule (like "how messy the cells look") is closely related to another (like "how deep the cells go"). If a doctor judges them separately, they might get confused or inconsistent.

This paper introduces a new computer program called INTERACT-CMIL that acts like a super-smart, consistent medical student to help grade these eye lesions.

The Problem: The "Five-Headed" Puzzle

The doctors need to answer five specific questions for every image:

WHO4: A specific classification level.
WHO5: A slightly different, more detailed classification level.
Horizontal Spread: How wide the lesion is.
Vertical Spread: How deep the lesion goes.
Atypia: How weird or "ugly" the cells look.

The tricky part is that these questions are interconnected. If the cells look very weird (Atypia), they usually also go deep (Vertical Spread). If a computer tries to answer these five questions five separate times, it might say, "The cells are weird" but also "The cells are shallow," which doesn't make sense in the real world.

The Solution: A Team of Experts Working Together

The authors built a system called INTERACT-CMIL. Think of it not as one giant brain, but as a team of five experts sitting around a table, all looking at the same picture.

Here is how their special training works:

1. The Shared Brain (Shared Feature Learning)

All five experts share the same "eyes." They look at the image through a powerful, pre-trained lens (called the CHIEF foundation model) that has already studied millions of other medical images. This gives them a strong foundation of what tissue looks like before they even start their specific jobs.

2. The "Three-at-a-Time" Game (Combinatorial Partial Supervision)

Usually, when you train a team, you ask everyone to work at once. But the authors found that if all five experts try to learn at the exact same time, they get confused or "overfit" (memorize the training data instead of learning the rules).

So, they invented a game: In every single practice round, only 3 out of the 5 experts are allowed to speak.

Round 1: Experts 1, 2, and 3 answer.
Round 2: Experts 2, 3, and 4 answer.
Round 3: Experts 1, 4, and 5 answer.

By cycling through every possible combination of three, the system forces the "Shared Brain" to learn features that are useful for any group of three, making the whole team much more robust and less likely to make mistakes.

3. The "Consistency Police" (Inter-Task Dependency Loss)

This is the most clever part. The system has a "Consistency Police" officer.
If the team predicts that the cells are "very weird" (High Atypia) but also "very shallow" (Low Vertical Spread), the Police officer says, "Wait a minute! In the real world, weird cells usually go deep. These answers don't match!"

The system uses a mathematical rule to punish the team if their answers contradict each other. This ensures that the final report is coherent. If the cells look dangerous, the whole report says they are dangerous, not just one part of it.

The Results: How Well Did They Do?

The team tested this new system on a collection of 486 real eye biopsy images from three different hospitals in Germany. They compared their "Team of Five" against two other methods:

A standard computer vision model (like a basic ResNet).
A standard model using the pre-trained "CHIEF" lens but without their special teamwork tricks.

The Winner: INTERACT-CMIL crushed the competition.

It improved the accuracy for the "WHO4" classification by 55% compared to the next best method.
It improved the "Vertical Spread" detection by 25%.

The paper claims that by forcing the computer to respect the relationships between the five different rules, it produces results that are much closer to what human experts would agree on.

Summary

Imagine trying to solve a complex puzzle where the pieces are connected. If you try to force the pieces together one by one without looking at the whole picture, they might not fit. INTERACT-CMIL is like a puzzle solver that:

Uses a giant library of knowledge (Foundation Model) to see the pieces clearly.
Practices by solving small groups of pieces at a time (Selective Supervision) to avoid getting stuck.
Constantly checks to make sure the pieces fit together logically (Inter-Task Consistency).

The result is a computer tool that can grade these tricky eye lesions more accurately and consistently than previous methods, offering a new standard for digital eye pathology.

Technical Summary: INTERACT-CMIL

Problem Statement
Accurate grading of Conjunctival Melanocytic Intraepithelial Lesions (CMIL) is critical for predicting conjunctival melanoma and determining treatment. However, current diagnostic practices rely on subjective histopathological examination of Hematoxylin and Eosin (H&E) stained sections, leading to significant inter-observer variability, particularly for low-grade lesions. The complexity of CMIL grading arises from the need to evaluate five interrelated histopathological axes simultaneously: WHO4 classification, WHO5 classification, horizontal spread (HS), vertical spread (VS), and cytologic atypia (AP). These criteria exhibit strong dependencies (e.g., severe atypia often coincides with increased vertical proliferation), yet existing computational approaches often model these axes independently, risking clinically implausible or inconsistent predictions. Furthermore, the rarity of CMIL and the scarcity of annotated data pose challenges for training robust deep learning models.

Methodology
The authors propose INTERACT-CMIL, a multi-head deep learning framework designed to jointly predict the five diagnostic axes while enforcing consistency across tasks. The architecture is built upon the CHIEF foundation model, a large-scale pathology encoder pretrained on diverse whole-slide images, which serves as a frozen feature extractor to capture transferable histomorphologic representations.

The core methodology consists of three key components:

Shared Feature Learning with Combinatorial Partial Supervision (SFCS): The framework utilizes a shared latent representation transformed into five task-specific classification heads. Instead of optimizing all heads simultaneously, the model employs a combinatorial partial supervision strategy. During each training iteration, only a subset of three heads is activated. The system systematically cycles through all possible three-head combinations, ensuring uniform exposure to diverse supervision contexts. This approach mitigates overfitting and encourages the shared representation to encode features that generalize across correlated but non-identical tasks.
Inter-Task Dependency Regularization: To address the structured correlations between grading axes, the authors introduce a divergence-based dependency loss ( $L_{dep}$ ). This loss function aligns the predicted joint task distributions with the empirical co-occurrence statistics observed in the data. By minimizing the Kullback-Leibler (KL) divergence between the predicted joint distribution and the empirical prior, the model enforces semantic coherence and consistency among the outputs of the different heads.
Overall Objective: The total loss function combines the classification loss ( $L_{cls}$ ) from the active heads and the dependency loss ( $L_{dep}$ ), weighted by a hyperparameter $\lambda$ . This ensures that the model optimizes for both task-level accuracy and cross-task consistency.

Key Contributions

First Specialized Framework: The paper presents the first deep learning framework specifically designed for CMIL classification, integrating both WHO and C-MIN grading paradigms.
Novel Architecture: It introduces a multi-head architecture utilizing combinatorial partial supervision to promote robustness and balanced learning across correlated tasks.
Dependency Modeling: The proposal of an inter-dependence loss that explicitly captures cross-criterion relationships, enhancing model interpretability and diagnostic coherence.
Dataset Curation: The creation and release of a newly curated, multi-center dataset comprising 486 expert-annotated conjunctival biopsy patches from three German university hospitals (Bonn, Tübingen, and Erlangen), providing a robust foundation for multi-task learning in this domain.

Results
The framework was evaluated on a held-out test set using 5-fold cross-validation, with performance measured by macro F1 scores. INTERACT-CMIL achieved consistent improvements over both a standard CNN baseline (ResNet-18) and a foundation-model baseline (BaseCHIEF) trained with single-task or joint multi-task objectives.

Performance Gains: Relative to the strongest baseline (BaseCHIEF), INTERACT-CMIL achieved relative macro F1 gains of 55.1% for WHO4, 24.7% for WHO5, 15.2% for Horizontal Spread, 25.0% for Vertical Spread, and 12.5% for Atypia.
Ablation Studies: The authors conducted ablation studies to quantify the contribution of each component:
- Removing the dependency loss caused the largest single performance drop (approx. 2–3% mean macro F1 decrease), particularly affecting atypia and vertical spread detection.
- Disabling temperature scaling further reduced performance by ~5–6%.
- Eliminating selective multi-head supervision led to the most significant cumulative drop (averaging over 8% below the full model), confirming its role as a strong regularizer that prevents gradient domination and balances learning across tasks.
Qualitative Analysis: Visual comparisons demonstrated close agreement between ground truth and model predictions, including in borderline cases where human variability is typically high.

Significance and Claims
The paper claims that INTERACT-CMIL establishes a reproducible, interpretable, and clinically aligned framework for automated CMIL classification. By explicitly modeling diagnostic dependencies through shared feature learning and dependency regularization, the system produces coherent multi-criteria predictions that align with expert consensus and established WHO/C-MIN standards. The authors position this work as a step toward standardized digital ocular pathology, offering a computational benchmark that addresses the subjectivity and variability inherent in current manual grading practices. The study emphasizes that the largest performance gains were observed in tasks relying on subtle morphological cues and inter-criterion consistency (vertical spread and cytologic atypia), areas where single-task or standard CNN-based models tend to fail.