# The Case for Local AI Development: Lessons From Computer‑Aided Detection of Tuberculosis and Silicosis in Southern Africa’s Ex‑Miners

## TL;DR

This paper argues for locally developed AI tools to detect lung diseases in Southern African miners, highlighting biases in existing systems and the need for local data and open practices.

## Contribution

The paper presents empirical evidence of AI biases in TB and silicosis detection and advocates for local public-sector AI development in underserved regions.

## Key findings

- Local chest X-rays from silica-exposed mineworkers form a distinct cluster separate from international datasets.
- Reducing image resolution disproportionately harms silicosis detection compared to TB detection.
- Accuracy metrics alone do not reflect clinical reliability, potentially hiding deployment failures.

## Abstract

The co‑epidemic of silicosis and tuberculosis (TB) in South Africa’s mining industry affects a large number of migrant workers and is compounded by limited access to chest X‑ray (CXR) screening. Although artificial intelligence (AI)‑based computer‑aided detection (CAD) systems for TB have demonstrated impressive accuracy against microbiological standards, validation among silica‑exposed populations has been limited. Moreover, well‑documented biases hinder CAD utility in diverse patient populations, potentially exacerbating existing healthcare inequities. In this article, we describe the challenges in developing CAD systems for TB and silicosis and present the potential benefits local public‑sector development initiatives can bring.

Using a local dataset of 2000 CXRs from silica‑exposed Southern African mineworkers, alongside publicly available international datasets and pretrained CAD models, we present empirical evidence of CAD biases. Dimensionality reduction analysis produced visual mappings that demonstrate how local CXRs form a distinct cluster, separate from international images. We also found that, relative to TB, reducing image resolution disproportionately degraded silicosis detection. Further visualizations proved that accuracy metrics alone are insufficient measures of clinical reliability, possibly obscuring deployment failures.

We conclude that local public‑sector CAD development offers a viable alternative to reliance on externally developed systems that likely exclude underserved populations. Addressing CAD deficiencies requires curating population‑representative datasets that capture local epidemiology and transparent, open‑source development practices that enable peer review and bias correction. Embedding technical and clinical expertise locally can transform AI‑based CAD from a potential instrument of digital colonialism into a mechanism that produces contextually appropriate diagnostics while advancing knowledge for equitable AI deployment worldwide.