An AI‑enabled, microscope‑integrated tool for decentralizing breast cancer diagnosis in Uganda: A mixed‑methods usability and workflow feasibility study (Preprint)

BACKGROUND Breast cancer is a leading cause of cancer morbidity and mortality worldwide, with disproportionately poor outcomes in low- and middle-income countries. In Uganda, delays between clinical presentation and diagnosis are driven by limited pathology capacity, long specimen transport times and weak referral systems. A task-shifting, low-cost diagnostic tool that brings cytology assessment closer to patients could reduce diagnostic delay and loss to follow-up. OBJECTIVE To evaluate and compare the usability and workflow feasibility of two Ekyaalo Diagnostics hardware form factors (V1 and V2), a low-cost AI-assisted platform for decentralized fine needle aspiration cytology slide digitization and remote pathology review across urban and rural health facilities in Uganda. It consists of a standard light microscope, off-the-shelf optical components and a smartphone application. METHODS We conducted a parallel, mixed-methods usability study to compare two hardware form factors across clinical settings in Uganda, in partnership with Makerere University, Mulago National Referral Hospital (NRH) and Mbarara Regional Referral Hospital (RRH). Study 1 evaluated V1 (smartphone adapter) at two tertiary hospitals (Mulago NRH and Mbarara RRH), while Study 2 evaluated V2 (eyepiece camera) across 11 rural and urban healthcare facilities spanning Health Centre IV–VII levels. Outcomes included System Usability Scale (SUS), ROI capture performance, Likert-scale usability measures and qualitative feedback on workflow integration and AI-assisted diagnosis. Non-parametric tests were used for group comparisons, and thematic analysis was used to analyze qualitative data from interviews and direct observations. RESULTS Both studies yielded encouraging results. In Study 1, mean SUS was 77.24 ± 17.18 indicated “good” usability and no statistically significant SUS differences across roles. Among 38 participants with capture data, mean regions of interest (ROI) captured was 6.39 ± 9.66, approximating routine review. AI outputs were viewed as supportive of the pathologist, rather than definitive. In Study 2, mean SUS was 66.59 ± 98.87, reflecting increased perceived complexity in rural contexts. Around 80% reported low confidence in acting on AI outputs without pathologist confirmation. Qualitative findings highlighted improved image quality and setup with V2 eyepiece camera but persistent constraints: shared microscopes, variable slide quality and a need for scanning the whole‑slide rather than sampling ROIs. CONCLUSIONS Ekyaalo Diagnostics demonstrated promising usability and strong acceptance among referral laboratory staff in Uganda. Ongoing iterative refinements that prioritize integration in rural settings, particularly by providing broader slide context and minimizing disruption to shared microscopes, are expected to enhance performance and acceptability. Given the similar challenges faced by low- and middle- income countries, Ekyaalo Diagnostics has the potential to reduce diagnostic delays in low-resource settings worldwide by supporting decentralized diagnostic workflows in low-resource settings.