The Lobular Breast Cancer Alliance (LBCA) recently posed questions to Tali Amir, MD, about her recent study A role for breast ultrasound Artificial Intelligence decision support in the evaluation of small invasive lobular carcinomas. Dr. Amir and senior author of the study, Victoria Mango, MD, are board-certified radiologists at Memorial Sloan Kettering Cancer Center. The study analyzes a new Artificial Intelligence (AI) system that can be used to assess small invasive lobular carcinoma (ILC) tumors seen using ultrasound. Dr. Amir answered LBCA’s questions about her research and the future of ILC care. Read her summary of the study and the other answers to the questions we posed below.
LBCA: Dr. Amir, what did your study aim to do and what do the results indicate?
Dr. Amir: Our study looked at the performance of an Artificial Intelligence Decision Support (AI DS) tool in the assessment of invasive lobular cancers seen on breast ultrasound. We tested the AI on ultrasound images of 83 invasive lobular cancers, and 100% were accurately characterized by the AI tool as warranting a biopsy. Furthermore, the AI DS tool reported a higher level of suspicion than the human reader (radiologist) when evaluating smaller lobular cancers. These results suggest that the AI DS tool used by a human reader may increase the radiologist’s confidence when evaluating smaller invasive lobular cancers on ultrasound.
LBCA: Can you explain how the AI DS function is added to the ultrasound? Is it software and/or hardware added to the ultrasound machine? Is it used in real time in conjunction with the ultrasound examination?
Dr. Amir: The AI DS is a software tool that can be used on the radiologist’s computers at the time of image interpretation, after ultrasound images are obtained, or while scanning a patient on the ultrasound machine. In our study, which was a retrospective design, we ran the AI DS analysis at a later date, and compared the findings of the analysis to the radiologist assessment given at the time of the patient’s exam.
In order for the AI DS to run an assessment, the radiologist must first detect the presence of a mass or area of concern on ultrasound. A box is drawn around the finding on the ultrasound image and then the AI system runs an analysis with the results of the analysis referred to as “output” providing degrees of likelihood of malignancy. The “outputs” from the AI DS analysis yield four levels of “likelihood of malignancy” of the tumor: “benign,” ” probably benign,” ” suspicious for malignancy,” and “probably malignant.”
LBCA: Can you explain in more detail how the test on the 75 women with ILC and 83 lesions was performed retrospectively? Was the test to corroborate the finding of ILC and the sizes and locations of the ILC tumors? Has the technology been tested on others with ILC?
Dr. Amir: We looked at 83 invasive lobular cancers diagnosed in 75 women between 2017 and 2019. For the purpose of the study, we ran the AI analysis retrospectively, not at the time of the patient’s imaging exams. For each cancer, using the AI software we drew a box around the mass and ran the AI DS tool. Outputs of “suspicious for malignancy” and “probably malignant” indicate that biopsy is warranted. We then compared the output of the AI DS with the assessment the radiologist gave at the time of the patient’s exam. We did not use the software to corroborate the size or location of the tumor, but instead the level of suspicion of the detected mass. To our knowledge, no additional studies have been performed specifically assessing the role of this AI DS in ILC.
LBCA: Are there limitations on what tumor size (how small or how large) can be detected?
Dr. Amir: In our experience, if the mass is large enough to be detected by the radiologist and not larger than the field of view of our ultrasound equipment, the AI DS tool can be used. Of note, the AI system does not detect the tumor; the radiologist must first identify the finding with ultrasound. Once a breast finding is identified in ultrasound, the AI DS helps determine the level of suspicion about malignancy and supports the recommendation regarding biopsy.
LBCA: Can you explain what interval cancer means and why this technology can decrease its incidence?
Dr. Amir: In breast imaging, an interval cancer means a cancer that is diagnosed within one year of a normal breast imaging exam, such as a normal mammogram and/or ultrasound. We can walk through an example. A woman may come in for her screening mammogram in January, and it is interpreted as normal. Then in June of that same year, she presents with a lump and her mammogram and/or ultrasound reveals cancer. She was diagnosed with cancer before she was due for her next screening mammogram, or in other words, in the interval between her screening mammograms.
In the case of interval cancers, it is impossible for us to know if the tumor cells were present at the time of the last study but were invisible to our eye on the mammogram or ultrasound (these are known as occult cancers) or if the tumor cells developed since the last exam. ILC tumor cells can blend into the surrounding normal breast tissue, making distinguishing the presence of tumor extremely difficult. At this time, the AI DS software we used relies on a radiologist to identify that an abnormality is in fact present. There is ongoing ultrasound AI research of technology that may help to point the radiologist to areas of interest for review, helping to find abnormalities that otherwise appear invisible and as a result, decrease interval cancers.
LBCA: Per your reported results, 100% accuracy in finding ILC when used with ultrasound, what are the proposed use(s) of this tool and at what point in the diagnosis and/or treatment process?
Dr. Amir: The AI DS accurately assessed all the ILCs as “suspicious for malignancy” or “probably malignant,” which also means that not a single ILC was incorrectly assessed as “probably benign” or “benign.” Our results suggest that the system is highly sensitive at assessing ILCs as suspicious when they are, particularly when small. This may play out in clinical practice in two ways. First, when a radiologist is evaluating a tiny mass, if the AI DS output indicates a tumor is “highly suspicious for malignancy,” it may help support the radiologist’s decision-making process leading to a recommendation for biopsy. This could result in more diagnoses when tumors are smaller in size. Second, and conversely, if the radiologist detects a mass that they assess as “probably benign,” and the AI DS classifies the mass as “benign’ or “probably benign,” the radiologist can feel increased confidence in deciding not to recommend biopsy.
LBCA: Can you explain how the AI DS technology improves the methods of ILC detection? Does it work with just the ultrasound?
Dr. Amir: The AI DS supports radiologist decision making in accurately recommending ILC for biopsy. (The AI DS does not function to and was not designed to detect the presence or absence of a mass.) This AI DS is used for ultrasound but other AI tools (not investigated in this study) exist for mammography and MRI.
LBCA: Your team utilized Breast Imaging Reporting and Data System (BI-RADS) to assess your findings in this study. Can you explain how the BI-RADS system works and how it describes possible malignancies?
Dr. Amir: A BI-RADS assessment is a numerical score assigned by the radiologist with the results of a mammogram, ultrasound, or MRI and is associated with a likelihood of malignancy and recommendation for next steps. A BI-RADS 1 and 2 assessment corresponds to a negative and benign result, respectively, or essentially a near 0% likelihood of malignancy. A B-RADS 3 assessment indicates a probably benign result, associated with a 2% or less likelihood of malignancy, and is usually followed closely with imaging. A BI-RADS 4 or 5 assessment indicates a result suspicious or highly suspicious for malignancy; BI-RADS 4 indicates a greater than 2% and up to 95% likelihood of malignancy while a BI-RADS 5 indicates a 95% or greater likelihood of cancer. BI-RADS 4 and 5 result in a biopsy recommendation.
LBCA: Was the AI DS technology used on any patients in addition to the 75?
Dr. Amir: In this study, the AI DS was used only on the ultrasound images of the 83 ILCs in the 75 patients described. Colleagues have investigated the use of this AI DS in additional patients. This software is FDA approved for use in clinical practice and is used at multiple medical centers and radiology practices.
LBCA: What do you perceive to be the significance of this study in the present world of lobular breast cancer diagnosis, and when do you think this technology will have an impact?
Dr. Amir: ILC presents a diagnostic challenge, in part due to its variable appearance on ultrasound. The AI DS tool proved to be extremely sensitive in assessing ILC as “suspicious’ or “probably malignant,” and particularly so for smaller tumors. Ultrasound features that radiologists assess to determine likelihood of malignancy, such as shape, margins, or presence of blood flow may be more difficult to discern in tiny masses, and therefore the utility of AI DS even greater. Use of the AI technology may help radiologists more accurately identify which breast ultrasound findings warrant biopsy.
LBCA: Will this technology be developed for uses with other breast cancer detection methods, such as mammography, or is it limited to ultrasound?
Dr. Amir: The AI DS is the product of a third-party company, so I cannot speak to future plans on their behalf. The software we used is specific for ultrasound. Alternate AI tools currently exist for mammography and MRI.
LBCA: You mention that the AI DS you used is an FDA approved device. Is it widely available now to radiologists?
Dr. Amir: The AI DS used in our study is a commercially available product that healthcare institutions may choose to purchase for clinical use in patient care. Research demonstrates encouraging results but further studies will help define how and in what context the AI DS is best utilized.
LBCA: One last question. Would you say that the import of your study was to demonstrate how using AI DS generally (regardless of the company that created it) in conjunction with ultrasound can help improve the detection of ILC tumors, particularly small ones OR to demonstrate that a particular type of AI DS has 100% accuracy?
Dr. Amir: While the results of ours study (that AI DS identified 100% of ILC lesions as suspicious) are based on the AI DS software we used, our study highlights that AI DS, in general, may help radiologists’ confidence when assessing ILC on ultrasound.