From Clusters to Classes: An Integrated Unsupervised and Supervised Deep Learning Approach for Explainable Chest X-ray Analysis
Faculty Mentor
Jongyeop Kim
Location
Russell Union Ballroom
Type of Research
Completed
Session Format
Poster Presentation
College
Allen E. Paulson College of Engineering & Computing
Department
Information Technology
Abstract
Automated chest X-ray analysis has advanced significantly through deep learning; however, most existing approaches depend heavily on large annotated datasets and lack interpretability, limiting their clinical adoption. This study presents an integrated unsupervised and supervised deep learning framework for explainable chest X-ray analysis that operates without expert-labeled data. Using a subset of the NIH Chest X-ray dataset (N = 1,609), we developed an end-to-end pipeline combining ResNet-18 feature extraction, autoencoder-based dimensionality reduction, and Deep Embedded Clustering (DEC) for autonomous pattern discovery.
The DEC model identified four well-separated clusters, supported by internal validation metrics (Silhouette Score = 0.4431; Davies-Bouldin Index = 0.9257). Supervised classifiers trained on the learned representations achieved up to 98.1% accuracy in predicting cluster membership, confirming strong cluster separability and feature robustness. Multi-level explainability was incorporated using SHAP for latent feature attribution and Grad-CAM for spatial visualization, revealing clinically interpretable radiographic patterns related to image brightness, contrast, and projection characteristics.
The proposed framework demonstrates how unsupervised deep learning, combined with explainable AI, can organize and interpret chest radiographs in a transparent and scalable manner. This work contributes toward the development of agentic, label-efficient clinical decision support systems capable of autonomous triage, workflow optimization, and trustworthy AI deployment in medical imaging environments
Program Description
.
Start Date
4-23-2026 2:00 PM
End Date
4-23-2026 4:00 PM
Recommended Citation
Coffie, Lord, "From Clusters to Classes: An Integrated Unsupervised and Supervised Deep Learning Approach for Explainable Chest X-ray Analysis" (2026). GS4 Student Scholars Symposium. 206.
https://digitalcommons.georgiasouthern.edu/research_symposium/2026/2026/206
From Clusters to Classes: An Integrated Unsupervised and Supervised Deep Learning Approach for Explainable Chest X-ray Analysis
Russell Union Ballroom
Automated chest X-ray analysis has advanced significantly through deep learning; however, most existing approaches depend heavily on large annotated datasets and lack interpretability, limiting their clinical adoption. This study presents an integrated unsupervised and supervised deep learning framework for explainable chest X-ray analysis that operates without expert-labeled data. Using a subset of the NIH Chest X-ray dataset (N = 1,609), we developed an end-to-end pipeline combining ResNet-18 feature extraction, autoencoder-based dimensionality reduction, and Deep Embedded Clustering (DEC) for autonomous pattern discovery.
The DEC model identified four well-separated clusters, supported by internal validation metrics (Silhouette Score = 0.4431; Davies-Bouldin Index = 0.9257). Supervised classifiers trained on the learned representations achieved up to 98.1% accuracy in predicting cluster membership, confirming strong cluster separability and feature robustness. Multi-level explainability was incorporated using SHAP for latent feature attribution and Grad-CAM for spatial visualization, revealing clinically interpretable radiographic patterns related to image brightness, contrast, and projection characteristics.
The proposed framework demonstrates how unsupervised deep learning, combined with explainable AI, can organize and interpret chest radiographs in a transparent and scalable manner. This work contributes toward the development of agentic, label-efficient clinical decision support systems capable of autonomous triage, workflow optimization, and trustworthy AI deployment in medical imaging environments
