Causality-Inspired Graph
Neural Network for
Interpretable Strabismus
Subtype Classification
Tien-Bach-Thanh Do
Network Science Lab
Dept. of Artificial Intelligence
The Catholic University of Korea
E-mail: osfa19730@catholic.ac.kr
2025/12/22
Jiawen Zheng et al.
MICCAI 2025
2
Introduction
● A common visual disorder (affecting ~3% of global population) where eyes are misaligned, potentially leading to
permanent vision impairment if not treated early
● Current challenges:
● 1: Existing DL methods lack interpretability
● 2: Most current systems only detect the presence of strabismus but fail to classify fine-grained subtypes
3
Proposed Method
• They propose the Causality-Inspired GNN:
• 1: Instead of relying on raw image pixels, they identify causally related visual features and structures them into
graph
• 2: Enhance diagnostic accuracy, ensure high efficiency for edge deployment
4
Proposed Method
Fig. 2. Overview of CI-GNN framework. The framework consists of 4 stages, including data acquisition, candidate feature extraction, causal feature selection, and
graph data modeling
5
Proposed Method
● Patients are instructed to gaze 9 specific directions by Hirschberg test, to capture abnormal eye movement
patterns
● Facial landmarks are detected to construct 9-gaze photograph in a bottom-to-top, left-to-right order
6
Proposed Method
● Interpretable model behavior requires extracting high-level features from raw image pixels, such as iris and orbital
centers
● Perform candidate feature extraction guided by clinical guidelines and ophthalmologists’ observations
● Construct separate candidate feature pools for each task
Candidate Feature Extraction
7
Proposed Method
● Constructed candidate feature pool contains both continuous variables (coordinate points) and discrete variables
(strabismus diagnostic categories such as direction and angle)
● To extract essential diagnostic features, employ score-based causal reasoning approach from HCM algorithm
● First apply PC algorithm and MRCIT for skeleton learning, followed by greedy algorithm for causal directed acyclic
graph construction
● Add edge based on maximum score gain for DAG
j-th feature empirical probability i-th observation
• At each iteration, the gain from adding potential edge l->j is computed
• To refine feature selection, extract Markov blanket of the target variable y (strabismus diagnosis outcome),
comprising its direct causes, direct effects, and variables that render y conditionally independent from the rest
Causal Feature Selection
8
Proposed Method
● Define graph structure as:
● V denotes 9 nodes corresponding to gaze positions in the 9-gaze photograph
● Graph is fully connected
● Apply GCN to extract features
Graph Data Modeling
9
Experiments
● Dataset:
o 1,075 real cases labeled with 6 categories
o 1: A-pattern
o 2: V-pattern
o 3: non-AV-pattern
o 4: esotropia
o 5: exotropia
o 6: vertical strabismus
10
Experiments
11
Experiments
12
Conclusion
● Proposed CI-GNN for interpretable strabismus subtype classification
● By integrating causal discovery with GNNs, they enhances both diagnostic accuracy and interpretability

251222_Thanh_LabSeminar[Causality-Inspired Graph Neural Network for Interpretable Strabismus Subtype Classification].pptx

  • 1.
    Causality-Inspired Graph Neural Networkfor Interpretable Strabismus Subtype Classification Tien-Bach-Thanh Do Network Science Lab Dept. of Artificial Intelligence The Catholic University of Korea E-mail: osfa19730@catholic.ac.kr 2025/12/22 Jiawen Zheng et al. MICCAI 2025
  • 2.
    2 Introduction ● A commonvisual disorder (affecting ~3% of global population) where eyes are misaligned, potentially leading to permanent vision impairment if not treated early ● Current challenges: ● 1: Existing DL methods lack interpretability ● 2: Most current systems only detect the presence of strabismus but fail to classify fine-grained subtypes
  • 3.
    3 Proposed Method • Theypropose the Causality-Inspired GNN: • 1: Instead of relying on raw image pixels, they identify causally related visual features and structures them into graph • 2: Enhance diagnostic accuracy, ensure high efficiency for edge deployment
  • 4.
    4 Proposed Method Fig. 2.Overview of CI-GNN framework. The framework consists of 4 stages, including data acquisition, candidate feature extraction, causal feature selection, and graph data modeling
  • 5.
    5 Proposed Method ● Patientsare instructed to gaze 9 specific directions by Hirschberg test, to capture abnormal eye movement patterns ● Facial landmarks are detected to construct 9-gaze photograph in a bottom-to-top, left-to-right order
  • 6.
    6 Proposed Method ● Interpretablemodel behavior requires extracting high-level features from raw image pixels, such as iris and orbital centers ● Perform candidate feature extraction guided by clinical guidelines and ophthalmologists’ observations ● Construct separate candidate feature pools for each task Candidate Feature Extraction
  • 7.
    7 Proposed Method ● Constructedcandidate feature pool contains both continuous variables (coordinate points) and discrete variables (strabismus diagnostic categories such as direction and angle) ● To extract essential diagnostic features, employ score-based causal reasoning approach from HCM algorithm ● First apply PC algorithm and MRCIT for skeleton learning, followed by greedy algorithm for causal directed acyclic graph construction ● Add edge based on maximum score gain for DAG j-th feature empirical probability i-th observation • At each iteration, the gain from adding potential edge l->j is computed • To refine feature selection, extract Markov blanket of the target variable y (strabismus diagnosis outcome), comprising its direct causes, direct effects, and variables that render y conditionally independent from the rest Causal Feature Selection
  • 8.
    8 Proposed Method ● Definegraph structure as: ● V denotes 9 nodes corresponding to gaze positions in the 9-gaze photograph ● Graph is fully connected ● Apply GCN to extract features Graph Data Modeling
  • 9.
    9 Experiments ● Dataset: o 1,075real cases labeled with 6 categories o 1: A-pattern o 2: V-pattern o 3: non-AV-pattern o 4: esotropia o 5: exotropia o 6: vertical strabismus
  • 10.
  • 11.
  • 12.
    12 Conclusion ● Proposed CI-GNNfor interpretable strabismus subtype classification ● By integrating causal discovery with GNNs, they enhances both diagnostic accuracy and interpretability