Enterprise AI Analysis

Ensemble Deep Learning for CXR Analysis

The proposed solution is an innovative ensemble deep learning framework designed for robust TB detection from chest radiographs. It integrates two distinct models:

• Convolutional Autoencoder (CAE): Focuses on unsupervised feature extraction, learning compact, high-level representations of CXRs. This enables the model to capture critical structural details without relying on labeled data in its initial training phase.
• Multi-Scale Convolutional Neural Network (MS-CNN) with Deep Layer Aggregation: Extends traditional ResNet architectures with multi-scale backbone modules and hierarchical layer connections. This design enhances feature representation by capturing complex patterns across various scales, ensuring computational efficiency.

By combining these models, the ensemble leverages complementary strengths, enhancing prediction robustness, reducing variance, and improving generalizability across diverse imaging conditions and patient populations.

Robust Validation Across Diverse Datasets

The framework was rigorously evaluated on three distinct datasets: two public (Montgomery County and Shenzhen from NLM/NIH) and one private clinical dataset from Songklanagarind Hospital, Thailand. Data preprocessing included DICOM loading, normalization, histogram equalization, and intelligent cropping to isolate lung fields and remove markings.

Performance was measured using key metrics such as AUROC, Accuracy, Sensitivity, Specificity, PPV, and NPV. The ensemble model consistently outperformed individual CAE and MS-CNN models, achieving a state-of-the-art AUROC of 0.98, with peak sensitivity and specificity of 99% and 94% respectively on the Shenzhen dataset. Expert radiologist review of misclassified cases confirmed the clinical relevance and diagnostic reliability, highlighting the model's ability to identify even atypical TB presentations.

Scalable AI for Global Health Equity

This automated TB screening tool offers significant strategic implications, particularly for low- and middle-income countries. It serves as a valuable triage tool for annual health check-ups and preoperative assessments, reducing diagnostic burden on limited radiological resources.

The solution's generalizability across diverse populations and imaging protocols makes it highly adaptable. Future work includes benchmarking against Vision Transformer (ViT) architectures, integrating bone suppression algorithms to enhance lesion visibility, and developing tailored models for HIV-associated TB, further expanding its impact. This AI system holds the potential to accelerate TB control efforts, leading to earlier diagnoses and ultimately saving countless lives.