💻 computer science

Towards Fair and Robust Volumetric CT Classification via KL-Regularised Group Distributionally Robust Optimisation

This paper proposes a KL-regularised Group Distributionally Robust Optimisation framework combined with a MobileViT-XXS and SliceTransformer architecture to simultaneously address distribution shifts across acquisition sites and demographic disparities in chest CT classification, achieving state-of-the-art performance on both COVID-19 detection and gender-fair lung pathology recognition tasks.

Original authors: Samuel Johnny, Blessed Guda, Frank Ebeledike, Goodness Obasi, Moise Busogi

Published 2026-03-18

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Samuel Johnny, Blessed Guda, Frank Ebeledike, Goodness Obasi, Moise Busogi

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

Imagine you are trying to teach a robot doctor to look at 3D CT scans (like high-tech X-rays) and diagnose diseases. You want this robot to be two things: smart (accurate) and fair (doesn't treat patients differently based on where they live or their gender).

This paper describes a new "teaching method" for this robot that solves two big problems that usually trip up AI in hospitals.

Here is the breakdown using simple analogies:

The Two Big Problems

The "Different Schools" Problem (Distribution Shift):
Imagine you teach a student using textbooks from School A. When you send that student to School B, they fail because School B uses different fonts, different paper, and different teachers.
- In the paper: CT scanners from different hospitals (different brands, different settings) produce images that look slightly different. An AI trained on Hospital A's images often gets confused when it sees images from Hospital B.
The "Classroom Majority" Problem (Demographic Fairness):
Imagine a teacher who only pays attention to the students who raise their hands the most. If a quiet student (a minority group) has a problem, the teacher might miss it because they are too focused on the loud majority.
- In the paper: In medical data, some groups are huge (e.g., "Male patients with Lung Cancer Type A") and some are tiny (e.g., "Female patients with Lung Cancer Type B"). Standard AI gets really good at the big groups but often fails the tiny, rare ones.

The Solution: A Smart, Balanced Teacher

The authors built a system with three main parts to fix these issues:

1. The "Slice" Strategy (The Architecture)

Instead of trying to swallow a whole 3D CT scan at once (which is like trying to eat a whole pizza in one bite and choking), the robot looks at the scan one "slice" at a time (like looking at individual pages of a book).

The Encoder: It uses a lightweight, efficient brain (called MobileViT) to read each slice.
The Aggregator: It uses a special "attention" mechanism (called SliceTransformer) to look at all the slices and say, "Okay, this specific slice has the most important clues, let's focus on that."
Why it matters: It's fast, cheap to run, and understands the 3D shape of the disease without needing a supercomputer.

2. The "Fairness" Training (Group DRO)

This is the core magic. Usually, AI learns by minimizing the average error. If it gets 100 easy cases right and 1 hard case wrong, the average looks great.

The Fix: The authors use Group Distributionally Robust Optimization (Group DRO).
The Analogy: Imagine a teacher grading a test. Instead of just looking at the class average, the teacher says, "I don't care about the average. I care about the student who got the lowest score." If a specific group (like "Females with a rare cancer") is doing poorly, the teacher forces the AI to spend extra time studying only those cases until they get better.

3. The "Stability" Brake (KL Regularization)

There was a catch with the "Fairness" method above. If you push the AI too hard to fix the worst group, it might panic and forget everyone else. It might start ignoring the majority groups completely to focus only on the one struggling group.

The Fix: They added a KL Regularization term.
The Analogy: Think of this as a "balance beam." The KL term is a gentle hand that says, "Hey, focus on the struggling student, but don't forget the rest of the class! Keep a healthy balance." It prevents the AI from going to extremes.

The Results: How Did They Do?

They tested this on two challenges:

Challenge 1: Detecting COVID-19 from 4 different hospitals.
- Result: Their method beat the previous best entry by a significant margin (5.9% better). It learned to handle the "different schools" problem well.
- Note: One hospital (Centre 2) was so different that even this smart AI struggled, which the authors admit is a hard limit of the current technology.
Challenge 2: Diagnosing 4 types of lung issues while being fair to men and women.
- Result: This is where they shined. They specifically targeted a tiny group: Females with Squamous Cell Carcinoma (only 5 examples in the training data!).
- The Win: Standard AI (Focal Loss) got this group right only 46% of the time. Their new method got it right 63.6% of the time.
- Overall: They improved the fairness score by 11% over the best previous entry.

The Bottom Line

This paper is about building an AI doctor that doesn't just get the "easy" answers right. It uses a clever balancing act to ensure it pays extra attention to the difficult cases (rare diseases) and the tricky environments (different hospitals), all while keeping the system small enough to run on standard hospital computers.

In one sentence: They taught an AI to be a fair and robust doctor by making it focus on its weakest students without forgetting the rest of the class.

1. Problem Statement

The paper addresses two critical challenges in deploying AI for chest Computed Tomography (CT) scans in clinical settings:

Distribution Shift (Domain Shift): Models trained on data from one hospital often fail to generalize to others due to differences in scanner hardware, reconstruction kernels, and imaging protocols.
Demographic Fairness: Standard training objectives optimize for average performance, often leading to poor results for underrepresented subgroups (e.g., specific gender-disease combinations like "Female Squamous cell carcinoma" with very few samples).

The authors tackle these issues simultaneously across two tasks:

Task 1: Binary COVID-19 classification across four different acquisition centers (multi-site domain shift).
Task 2: Four-class lung pathology recognition (Adenocarcinoma, Squamous cell carcinoma, COVID-19, Healthy) with explicit fairness constraints based on gender.

2. Methodology

A. Model Architecture

The framework is designed to be lightweight (under 2M parameters) to prevent overfitting on small medical datasets while capturing volumetric context.

Slice Encoder: Uses a shared MobileViT-XXS backbone to encode individual 2D axial slices independently. The model is adapted for single-channel CT input by averaging pretrained RGB weights.
Volumetric Aggregator: A SliceTransformer (a two-layer Transformer Encoder) aggregates the slice embeddings. Unlike simple pooling, this uses self-attention to learn which anatomical slices carry the strongest diagnostic signals, preserving inter-slice spatial context.
Classification Head: A linear classifier with LayerNorm and Dropout produces the final logits.

B. Training Objective: KL-Regularised Group DRO

The core innovation is a modified Group Distributionally Robust Optimisation (Group DRO) objective.

Standard Group DRO: Minimizes the worst-case loss over predefined groups by adaptively upweighting groups with high loss. However, on small datasets, this can lead to weight collapse, where the model focuses exclusively on a single "hard" group, ignoring others.
KL Regularisation: The authors introduce a Kullback-Leibler (KL) divergence penalty term ( $\alpha \cdot D_{KL}(w \parallel u)$ $α \cdot D_{K L} (w ∥ u)$ ) to the loss function.
- This penalty encourages group weights ( $w$ ) to stay close to a uniform distribution ( $u$ ).
- Effect: It prevents the weights from collapsing onto a single group, providing a stable balance between protecting the worst-case subgroup and maintaining average performance.
Group Definitions:
- Task 1: Groups are defined by Acquisition Center (4 groups).
- Task 2: Groups are defined by Gender $\times$ Class (8 groups, e.g., Male-Adenocarcinoma, Female-Squamous). This fine-grained definition is crucial for targeting severely underrepresented combinations (e.g., only 5 training samples for Female Squamous).

3. Key Contributions

Lightweight Volumetric Framework: A hybrid architecture combining MobileViT-XXS and SliceTransformer that achieves competitive performance with <2M parameters, suitable for small, multi-center medical datasets.
KL-Regularised Group DRO: A novel training objective that adaptively reweights acquisition centers and fine-grained demographic subgroups. The KL term specifically addresses the instability of standard Group DRO on small datasets.
Systematic Ablation: Comprehensive analysis of the KL regularisation strength ( $\alpha$ ) and aggregator design, offering practical guidelines for handling imbalanced medical data.

4. Experimental Results

Task 1: Multi-Source COVID-19 Detection

Performance: The best configuration ( $\alpha=0.5$ ) achieved an overall F1 score of 0.835.
Comparison: This outperforms the best published challenge entry (FDVTS, 0.776) by +5.9 percentage points and the weighted cross-entropy baseline by +3.1 points.
Observation: While the method improved performance across most centers, it failed to recover performance on Centre 2 (near-zero F1). The authors note this highlights a limitation: reweighting cannot compensate for a domain that is completely absent or radically different from the training distribution.

Task 2: Fair Disease Diagnosis

Performance: The best configuration ( $\alpha=0.5$ ) achieved a mean per-gender macro F1 of 0.815.
Comparison: This surpasses the best challenge entry (70.4%) by +11.1 points and the Focal Loss baseline by +3.8 points.
Fairness Impact:
- The method significantly improved the Female Squamous cell carcinoma F1 score from 0.462 (Focal Loss) to 0.636 (GDRO $\alpha=0.5$ ).
- It reduced the performance gap between male and female subgroups to a minimum ( $\Delta=0.013$ ) at $\alpha=0.5$ .
- Ablation Insight: Without KL regularization ( $\alpha=0$ ), performance was good but not optimal. At $\alpha=1.0$ (forcing uniform weights), the model over-regularized, benefiting the majority group (Male) while hurting the minority (Female), proving that the KL term must be tuned carefully to balance protection and fairness.

5. Significance and Conclusion

This work demonstrates that KL-regularised Group DRO is a robust mechanism for addressing both domain shift and demographic bias in medical imaging, particularly when data is scarce and imbalanced.

Practicality: The lightweight architecture makes the approach feasible for deployment in resource-constrained environments (e.g., edge devices, portable scanners).
Fairness: By defining groups at the Gender $\times$ Class granularity, the method successfully targets specific, severely underrepresented patient populations that standard fairness methods often miss.
Limitations & Future Work: The failure on the unseen domain (Centre 2) suggests that while reweighting helps with known shifts, future work must incorporate test-time adaptation or site-specific normalization to handle completely novel acquisition protocols.

In summary, the paper provides a principled, effective, and computationally efficient solution for building fair and robust AI models for volumetric CT diagnosis.