The primary aim of this study was to assess the accuracy of trained technicians in performing liver segmentation, Couinaud segmentation and lesion detection and delineation. The experienced technicians' performance was compared with that of experienced radiologists, with a particular focus on inter-operator variability. The results showed excellent agreement between technicians and radiologists.

The participating radiologists had an average of 13 years of experience, while the technicians had backgrounds in biology but no prior experience in radiology. Despite this, the liver volumetry results demonstrate that, with appropriate training and AI assistance, technicians are capable of accurately delineating livers and partitioning the volume into Couinaud segments based on manually placed anatomical landmarks. Significantly, their results agreed strongly with those of the radiologists, with inter-rater variability similar to that observed between radiologists. With respect to lesion detection, all lesions segmented by at least two out of three radiologists were also identified by most technicians. However, technicians detected an additional 15 structures that were not flagged by radiologists. An adjudicator panel, composed of two radiologists with 15 and 18 years of experience, assessed these structures and agreed that 7 out of these 15 identified structures were indeed lesions; the remaining 8 structures were deemed to be non-lesion, representing inconsequential, benign findings such as partial volume artifacts or blood vessels (Example in Fig. 6). Among the lesions delineated by most participants in both groups, there was remarkable consistency in terms of diameter and Couinaud location, indicating large independence of the analyst, minimizing the potential for variations that could affect treatment decisions and planning based on lesion characteristics. Discrepancies between the groups were lower than those measured within each group, primarily because accuracy samples are based on the mean of three measurements, whereas group variabilities compare single measurements. Variability within both groups was comparable for lesion diameter categorization and location in Couinaud segments. Despite similar overall accuracies in lesion detection, the technicians group exhibited a tendency towards over-detection compared to radiologists, who had a higher specificity. Their agreement was partly aided by the initial mask provided by the AI model, which has been shown to increase agreement [14]. To further refine the pipeline for clinical deployment, a final review step by a radiologist could be introduced, focusing on removing false-positive lesions flagged by technicians.

Example of structures segmented by technicians but not by radiologists. a Adjudicated by panel as lesion. b Adjudicated by panel as not-lesion

To address the challenges of radiology in recent times, new clinical strategies are being explored, such as radiologists overseeing pre-calculated outputs and employing personnel with lower levels of specialization, while integrating AI-based tools. Several studies have highlighted the potential of integrating AI with non-specialist personnel in radiology. For instance, Sullivan et al. demonstrated that technologists could effectively perform total metabolic tumor volume measurements with AI assistance [16]. Similarly, Suman et al. (2020) found that trained technologists could reliably conduct volumetric pancreas segmentation on CT images [17]. In the field of RECIST measurements, Gouel et al. showed high reproducibility when technologists performed RECIST 1.1 measurements in breast cancer follow-up [18], supported by Beaumont et al., who optimized workflows for RECIST assessments [19]. Studies have also shown that non-specialists can accurately quantify arterial obstruction in pulmonary embolism [20] and measure leg length discrepancies [21]. AI integration has significantly improved lesion detection accuracy in various contexts, including real-time skin lesion assessment, colonoscopy for diminutive polyp identification, and breast cancer detection in mammography [22,23,24]. These advancements underscore the potential for AI to support less specialized personnel in complex diagnostic tasks. Building on this foundation, our study demonstrates high agreement levels between radiologists and technicians in liver volumetry and lesion detection during liver resection. This suggests that AI-assisted non-specialist personnel can effectively reduce the workload on radiologists without compromising diagnostic accuracy. By leveraging AI tools, radiology departments can potentially optimize workflow efficiency and improve diagnostic outcomes.

The integration of AI into the workflow potentially bridges a proficiency gap, ensuring high-quality analysis and diagnosis even when highly skilled radiologists are not directly involved. By leveraging the strengths of AI and the human element provided by technicians, a more efficient and accessible diagnostic process is facilitated, optimizing the utilization of available radiological expertise. Furthermore, implementing an AI-assisted technician approach in radiology departments could potentially lead to cost savings. By reducing the reliance on highly specialized radiologists for routine tasks and improving workflow efficiency, this approach could enable radiology departments to handle a higher volume of cases without compromising quality.

This study has several limitations. First, the sample size of 18 cases is relatively small, and the range of liver pathologies analyzed was limited, focusing primarily on patients with colorectal liver metastases. A larger dataset with a wider variety of clinical scenarios, such as primary liver tumors and other metastatic origins, would strengthen the assessment of the AI-assisted technician approach and improve the generalizability of the findings. Second, the image dataset was confined to one scanner model and manufacturer, though our experience suggests this is not a severe limitation. Third, as the study aims to assist in treatment decision-making and planning for patients previously diagnosed with liver cancer, the operators were aware of the presence of at least one lesion. Providing diagnostic clinical information prior to the execution of the pipeline could further improve the specificity of the operators. Finally, a comparison between radiologists with and without AI assistance would have been valuable to evaluate the effect of the assistance and potential time reduction. Addressing these limitations in future studies would provide a more comprehensive evaluation of the AI-assisted technician approach in liver cancer imaging.

In conclusion, this study suggests that AI-assisted technicians can achieve performance comparable to radiologists in liver segmentation, Couinaud division, and lesion detection. While this approach may improve radiology department efficiency, it is important to consider the AI results alongside patient clinical information. Radiologist supervision remains essential for interpreting the AI output within the broader clinical context and making final patient management decisions. Future studies should explore the integration of AI-assisted technicians into clinical workflows, assessing both quantitative performance and impact on patient outcomes.