We used published data on efficiency gains in nodule detection by DL-CAD to calculate break-even points for various pricing models of commercial CAD systems. We performed the analysis in three representative Western countries, the USA, the UK, and Poland, in which lung cancer screening is implemented or currently considered [7,8,9].

A scoping review was performed to get an overview of reading time with and without DL-CAD assistance. The published studies were retrieved from the PubMed database from its inception to Oct 16, 2022. The search string consisted of the following keywords (Table S1): computed tomography, reading time, computer-aided detection, and lung nodule. Since we focused on the use of commercially available DL-CAD in a lung screening setting, articles were excluded if (1) contrast-enhanced CT was included, (2) non-deep learning-based CAD was evaluated, and (3) non-commercial DL-CAD was evaluated. The extracted data included the first author, publication year, country, normal or low dose CT, name of the commercial DL-CAD system, and reading time per CT scan with and without DL-CAD assistance. The reading time extracted from each study was pooled with random effect model.

Since no studies evaluated the reduction in reading time of DL-CAD as a pre-screening reader, we estimated it based on the proportion of normal CT scans in a screening setting. According to a recent review, 22–51% of participants (depending on the detection limit of size) in screening RCTs have a lung nodule detected at baseline [4]. We assumed that a DL-CAD would be able to exclude 80% of the nodule-free cases to allow the DL-CAD to be set to a very high sensitivity at a modest specificity just to be sure not to lose any potentially actionable nodules. Therefore DL-CAD as a pre-screening reader would reduce 39–62% of the workload of a radiologist and thus would save 39–62% reading time. We took the best-case and worst-case scenarios in which 62% and 39% of the reading time was saved, respectively.

Actual pricing for the DL-CAD system depends on country, pricing model, and local negotiations. While exact numbers are not publicly available, the National Institute for Health and Care Excellence (NICE) has published numbers for the UK in 2021 [10]. There are usually direct cost involved around installation and training, which were in the range of £8500–9000 (€9971–10,558, 1 GBP = 1.1731 EUR on average in 2022 [11]). In addition, there are yearly fees of £4000–14,800 (€4692–17,362) consisting of cost for hosting, monitoring, and support. Currently, three different pricing models are offered by various companies to pay for the actual use of the DL-CAD: a pay-per use model, a one-off perpetual software license, and a yearly subscription model. Data published by NICE suggests pay-per-use pricing of £5–7.5 (€5.9–8.8) and one-off pricing of £44,000 (€51,616). Yearly subscription costs usually vary according to the expected number of scans to be processed per year, estimated at €20,000 per year.

The payment to a radiologist per hour [12] is approximately £108 in the UK (€127, 1 GBP = 1.1731 EUR), $206 in the USA (€196, 1 USD = 0.9518 EUR [13]), and zł 211 in Poland (€45, 1 PLN = 0.2135 EUR [14]).

For these data, we calculated the break-even points at which using DL-CAD for nodule detection would start to become cost-saving using the three various pricing models, and the three reading paradigms. For the purpose of this study, we used the cost settings of the UK, USA, and Poland.

For the pay-per-use model, we calculated the cost per case for break-even. For the one-off perpetual software license and for the yearly subscription model we determined the break-even by calculating the minimum workload required to earn back the investment.

The cost (C) per case for break-even was calculated as C = S*Δt, where S are the gross salary costs of a radiologist in euros per hour, and Δt is the saved time in hours per CT scan with DL-CAD assistance. The minimum workload (W) in the number of assisting CT scans reading of a DL-CAD system was calculated as W = P/(S*Δt), where P is the price (in euros) of a commercial DL-CAD system, S are the gross salary costs of a radiologist in euros per hour, and Δt is the saved time in hours per case with DL-CAD assistance.