**3.2 How can AI transform the work of a radiologist?**

AI can play a transformative role to help unlock the solutions to many challenges of radiology such as increasing workload and staff shortages. AI can also transform the work of a radiologist by mainly following steps in image analysis, which includes detection, characterization, and monitoring in several ways [5, 11, 13] (see **Figure 2**). One of the most important areas where AI is used in radiology is **image analysis,** where AI-based systems can be trained to detect and identify specific structures or abnormalities on medical images such as tumors, blood vessels, or organ abnormalities. This can improve diagnostic accuracy and productivity by highlighting potential abnormalities that may have been missed by the human eye [5, 11].

Another important area is **characterization,** where AI-based systems can be trained to classify and characterize different types of abnormalities or lesions [5, 11]. For example, AI algorithms can be used to differentiate benign from malignant tumors or to classify different types of liver lesions. This can help radiologists to make more accurate and specific diagnoses and guide treatment decisions [5, 8]. This can also reduce the need for additional imaging or biopsies, which can save time and money.

**Figure 2.** *AI applications in radiology.*

**Monitoring** is another area where AI is being used in radiology, where AI-based systems can be used to monitor changes in lesions over time [5]. For example, AI-based systems can track the growth of a tumor or the response to treatment [11], which can help radiologists to make more informed decisions about patient care. This can also help to identify patients who need additional monitoring or treatment and can lead to improved patient outcomes.

AI is also being used in **image acquisition** by deep learning-based reconstruction algorithms that can reduce scan time and improve image quality, especially in MR imaging. MR imaging can take anywhere from 30 and 60 minutes and occasionally longer depending on the protocol. Some patients, particularly elderly patients, can become uncomfortable and claustrophobic lying in a confined space for this period of time. Being able to obtain high-quality imaging in a shorter time can help alleviate this issue and reduce the presence of motion artifact. This can ultimately improve the diagnostic confidence in the images and prevent unnecessary repeating sequences [14].

Additionally, AI can assist radiologists by providing them with **automated reports and summaries** that include relevant information and analysis, which can save time, reduce the workload and errors caused by manual reporting, and improve communication with other healthcare providers [8]. AI can also support radiologists by integrating with other healthcare systems, providing them with comprehensive patient information and data from other sources such as electronic health records, lab results, and previous imaging studies, which can provide a more comprehensive view of the patient's condition and assist in the diagnostic process.

It is important to note that AI in radiology still requires human interpretation and oversight, as AI algorithms are not perfect, they can make errors or miss certain findings. It is anticipated that AI in radiology will become increasingly more precise and reliable over time as more data is acquired and technology advances.

#### **3.3 Challenges of AI and liabilities around wrong and missed diagnosis**

Though AI has the promise to improve diagnostic accuracy and efficiency, it also poses certain liabilities related to wrong and missed diagnoses [9]. One potential liability is that AI systems may produce incorrect or unreliable results due to factors such as poor image quality, incorrect data input, or errors in the algorithms used to analyze the images [15]. This could lead to wrong, delayed, or missed diagnoses, or unnecessary treatments [9].

Another potential liability is that AI systems may not be able to detect certain types of lesions or diseases, particularly those that are rare or atypical. Additionally, there are concerns about the lack of standardized benchmarks to compare and validate AI models for practical implementation. This could lead to missed diagnoses, which can be particularly dangerous if the condition is serious or life-threatening.

AI systems also require proper validation and testing before they are used in clinical practice. Validating data sets is time-consuming, and thus can jam many machinelearning projects. Handling unexpected inputs such as artifacts and poor imaging can also pose a problem in high-quality data sets. Medical research on data sets can also act as a hurdle as many patients value their privacy [16, 17]. If they are not appropriately validated, they may not be suitable for the intended use or population, and this can lead to wrong or missed diagnoses as well.

Additionally, there is also a lack of reasoning and an inability to explain AI models [17]. There is the potential that AI systems could be used to override the judgment of

#### **Figure 3.**

*Efficacy and safety of images.*

radiologists, leading to an increased risk of wrong or missed diagnoses if the radiologist's judgment is ignored (see **Figure 3**).

To mitigate these liabilities, it is imperative to ensure that AI systems are properly validated and tested before they are used in clinical practice and that radiologists are properly trained in their use. Since investigating liability issues will require different skills from lawyers and additional evidence from technology along with medical expert opinion, we will need support from our technology law colleagues to design regulations [18]. It is vital to have proper governance, policies, and regulations in place for the use of AI in radiology.
