Artificial intelligence is widely expected to ease pressure on radiology services by supporting detection, prioritisation and reporting. A rapid evaluation examined how AI for chest diagnostics, including lung cancer, was procured and prepared for deployment across National Health Service (NHS) imaging networks in England. The work covered 12 networks involving 66 acute NHS Trusts and assessed the first phase of activity between March and September 2024. Using interviews, observations and document review, and guided by the Non-adoption, Abandonment, Scale-up, Spread and Sustainability (NASSS) framework, the evaluation identified where timelines slipped, which tasks consumed most effort and which contextual factors helped or hindered progress. The findings highlight the social and technical work required to implement AI at scale and underscore the importance of capacity, governance and engagement.
Procurement Took Longer and Demanded Wider Expertise
The Artificial Intelligence Diagnostic Fund (AIDF) launched in July 2023 to accelerate chest diagnostics through network-level procurement. Networks convened panels to set specifications, score tenders and select suppliers. This demanded diverse expertise spanning clinical pathways, radiology practice, information governance, local diagnostics infrastructure and procurement. Mature networks with established relationships were better placed to recruit the right mix of panel members. Nevertheless, some participants lacked confidence in differentiating between tenders, especially when submissions were lengthy, highly technical or insufficiently tailored to local systems. Several networks mitigated this by assigning reviewers to criteria aligned with their expertise and by sharing templates and learning across the programme.
Must Read: Safe Use of Generative AI in Radiology
Selection criteria covered service quality, costs and equity, diversity and inclusion. Most chosen tools scored highest on quality. Some procurements were complicated by clinicians’ prior relationships with vendors, reinforcing the value of independent scoring against agreed specifications. Following approval from national leadership, local contracting began but was frequently slowed by challenges raised by unsuccessful suppliers, including requests for feedback, disputes about scoring and, in some cases, legal concerns. Autonomy across Trusts led to variable contracting routes, from a single lead Trust model to multiple bilateral agreements, adding weeks to the timeline.
Timelines consistently slipped. Contracts originally expected by November 2023 were signed later in the in-depth cases, in March, May and September 2024. Clinical deployment, initially targeted for December 2023, began in some sites in May 2024. By November 2024, tools were operational in clinical services in 24 of 66 Trusts, rising to 43 by June 2025. Across the networks studied, only two of sixteen tendering suppliers were selected. Selected tools covered different modalities and functions, including chest X-ray or CT, triage of urgent cases and identification of potential symptoms. Use was as decision support, with human readers completing their assessment before reviewing AI outputs.
Integration and Governance Varied Across Trusts
Preparation for deployment centred on integrating tools with local Radiology Information Systems (RIS) and Picture Archiving and Communication Systems (PACS), updating workflows and obtaining permissions. Heterogeneity in RIS and PACS across Trusts meant integration often had to be repeated site by site. One network had already standardised systems, enabling shared development of worklist logic and smoother rollout. Another used the programme to move towards a common PACS to support future AI, though this increased near-term complexity and delay.
Project management and clinical leadership were critical enablers. Regular meetings between Trust teams and suppliers set milestones, tracked risks and coordinated testing on retrospective data, shadow mode operation and standard operating procedure updates. Close involvement of radiology leads, PACS managers and vendor engineers supported progress, especially where suppliers were new to NHS platforms and processes.
Information governance and patient safety approvals were necessary but inconsistent in pace and form across organisations. Different meeting schedules, templates and documentation standards limited reuse of materials across Trusts and increased workload. In some cases approvals were expedited by convening decision-makers within project meetings, such as clinical safety officers reviewing hazard assessments in real time. Where dedicated project managers were in place, coordination, engagement and timeliness improved, in their absence, tasks fell to clinical or network staff with limited capacity.
Engagement, Patient Communication and Data Needs
Staff engagement and training focused on ensuring appropriate use, reinforcing that AI serves as decision support and that human interpretation prevails. Many networks identified super users to receive vendor-led training and cascade it locally. Training covered interface features such as highlighted regions of interest, confidence indicators and documentation of concordance. Sites also considered how to assure training coverage for new and existing staff, with one hazard assessment prompting formal training logs. Some senior clinicians voiced concerns about autonomy, accountability and missed findings. Early, ongoing engagement was seen as necessary to address such concerns, but not all training addressed them directly.
Approaches to informing patients and carers varied and, at the time of data collection, were often still being finalised. Plans ranged from posters, leaflets and social or local media to direct communication, while some networks planned to inform only on request or not at all. Discussions balanced the potential of presenting AI implementation as positive for service efficiency with the need for transparency, particularly as patients may access records that indicate AI involvement. There was no general requirement to disclose use beyond existing rules, given classification as medical devices and the stipulation that tools should not influence clinical decisions.
Programme monitoring required networks and Trusts to provide benefits metrics at baseline and at six-monthly intervals post-deployment. Many struggled to assemble these data quickly due to complex reporting specifications and the need to link information across multiple systems not designed for automated extraction. Additional analytical support and network-level data platforms helped where available, but the effort demanded extra time and resources during early rollout.
Early procurement and deployment of chest AI across NHS imaging networks demonstrate that large-scale implementation is a complex socio-technical undertaking that exceeds initial timelines and demands sustained coordination. Diverse expertise, robust project management, consistent governance pathways and proactive engagement of clinicians, patients and suppliers were central to progress. Variation in IT infrastructure and data capability shaped local trajectories, while monitoring requirements added workload but supported evaluation. The experience indicates that AI can contribute to service delivery, yet realising benefits depends on adequate time, capacity and stakeholder engagement rather than rapid, uniform rollout.
Source: The Lancet Discovery Science - eClinicalMedicine
Image Credit: iStock
