Exploring Artificial Intelligence Based Interpretation of Medical Images in Radiological Workflows in Zambia

The manual nature of radiological workflows is a major issue affecting patient outcomes and the quality of healthcare provision in Zambia with only 9 radiologists in 2021. Artificial intelligence (AI) solutions are increasingly becoming mainstream. However the implementation and use of these solutions in low resource settings has been very slow. This study aimed to: (1) empirically characterize the workflow challenges regarding workload and effectiveness in Zambian radiology; (2) design and implement a locally adaptable, semiautomated AI solution for pneumonia detection; and (3) evaluate the impact of this solution on practitioner workload and user acceptance. Guided by the Cultural-Historical Activity Theory (CHAT) framework, a multi-phase mixed-methods design was employed. Phase I utilized a cross-sectional survey (n=100) and semi-structured interviews to assess baseline constraints. Phase II involved the development of binary classification (VGG16) and object detection (Mask R-CNN) models integrated into a web-based DICOM viewer. Phase III consisted of a controlled, within-subjects experiment (N=47) comparing manual versus AI-assisted interpretation using the NASA Task Load Index (NASA-TLX) and Technology Acceptance Model (TAM 2). Contextual inquiry revealed significant systemic breakdowns, with 74.5% of clinicians reporting reduced effectiveness due to missing clinical history and a heavy reliance on non-dedicated viewing devices. The developed AI models achieved a high Recall of 79%, prioritizing sensitivity to serve as a safety net for screening. Experimental results indicated a reduction in perceived workload, specifically in Mental Demand (Raw Score Δ-0.68), and high user acceptance, with a mean Perceived Usefulness of 5.7/7. The study concludes that humanin-the-loop AI effectively mediates the tension between high patient volumes and limited expertise by functioning as a cognitive scaffold for generalist doctors. It is recommended that health stakeholders adopt AI-enabled task shifting as a national strategy, provided it is accompanied by investments in digital infrastructure and mandatory user competence training to ensure safe deployment. IV