Why Radiologists Should Lead on CT Colonography
Early-onset colorectal cancer is rising; expanding access to CT colonography can boost screening, ease capacity strains, and improve early detection.
Read more


With the recent rapid advancement of artificial intelligence (AI) technology, its integration into the field of Radiology has become more prominent through its incorporation into various imaging techniques such as magnetic resonance imaging (MRI). As breast cancer is one of the most common cancers with a significant impact on mortality and global disease burden, early detection and treatment are crucial for optimal health outcomes. Efforts to use computer technology to aid breast cancer detection have taken place over 20 years ago with diminishing returns. Although screening mammographs are typically used to detect the early stages of breast cancer, MRI is useful for supplementary screening to confirm negative results. This is especially evident in interval cancers, which become symptomatic following a negative initial screen and before the following annual screen.
Annual supplemental breast cancer screening is undergone by over 500,000 women in the United States. Hirsch et al. hypothesized that AI could predict occurrence of breast cancer based on information contained in benign MRIs given that these MRIs contain details that can be used to determine the outcome of the next annual screen. This study used a pre-trained 2D convolutional neural network previously trained on patients at the clinical site the study took place. The design of this neural network enabled it to detect breast cancer in current MRIs through the assignment of a probability of each 2D sagittal slice containing cancer. Its overall output described the maximum probability across all slices of a particular breast. The network was optimized through the previous patient data. It was found that 10% of cases reevaluated by AI captured 30% of breast cancers diagnosed in the subsequent screening exam. This study was hindered by the limited number of cancers that can be detected by screening as well as the use of sagittal scans from a single clinical site.
AI has also been implemented in the process of determining a patient’s need for MRI screening post mammography. Salim et al. conducted ScreenTrustMRI, a randomized clinical trial involving an AI tool that scanned mammograms to determine need for supplementary MRI screening. The AISmartDensity software was used to screen each mammogram. This software consisted of three AI models capable of using an amalgamation of neural network approaches to classify images. These approaches included determination of inherent breast cancer risk through an EfficientNetB3 architecture, masking potential through ResNet34 architecture assigning normal mammograms to difficulty levels through individual radiologist assessments, and a ResNet34 architecture capable of detecting cancer signs. The software averaged standardized scores from these models to provide its assessment, and these scores were age-adjusted. For those participants whose supplemental MRIs were flagged as BI-RADS 3-5, further workup including biopsy was undergone. A cancer detection rate of 64.4 cancers for 1,000 MRI examinations determined necessary per AI was discovered with a 95% confidence interval. Further, a positive predictive value of 38% for participants indicated post MRI and 50.7% for biopsied participants was noted. Individuals that were determined to be classified as BI-RADS 3,4, and 5 have a PPV of 13%, 63%, and 85.7%, respectively. To date, AISmartDensity has not been implemented into regular clinical practice.
The use of AI in the field of Radiology warrants discussion relating to legality and ethics. As there is potential for errors with AI models, there is uncertainty surrounding legal liability when using these models autonomously. This becomes paramount in the instance of an incorrect diagnosis constituting grounds for medical malpractice if responsibility is placed on the radiologist instead of the software itself. Concerns regarding privacy must also be accounted for. Security measures must be put into place when AI comes into contact with large amounts of patient data in the event of a data breach. Data mining is a concern even if the data does not contain information directly identifying patients. Further, the potential of misuse of data can cause a sense of mistrust and discomfort by patients. AI models have the potential to become exposed to data containing bias, leading to unequal outputs.
It is worth noting that research regarding artificial intelligence primarily focuses on algorithm performance over clinically relevant outcomes. For example, MRI risk prediction studies do not tend to account for implications of clinical workflow. Other studies provide the concordance index of AI models, a statistic that relays how well the model is in predicting a series of events. Further studies shall be undergone to analyze the clinical implications of early detection through these models, especially as considerable effort is undergone in the reevaluation of MRIs by the radiologist.
Why Radiologists Should Lead on CT Colonography
Early-onset colorectal cancer is rising; expanding access to CT colonography can boost screening, ease capacity strains, and improve early detection.
Read more
Can and Cannot of AI
AI promises major efficiency gains in radiology but raises uncertainty about autonomy, liability and how much humans will remain essential.
Read more
Protect Your Energy, Water and Patients
Radiologists can boost sustainability by optimizing workflows, preparing for outages, and making smarter AI and energy choices in their practices.
Read more