Hear from Dr. Manisha Bahl about her award winning lobular study proposal, “Improving Detection of ILC on Digital Breast Tomosynthesis and Reducing False-Negative Rates with Artificial Intelligence.”
Dr. Bahl is an Associate Professor of Radiology at Harvard Medical School and Breast Imaging Division Quality Director and Co-Service Chief at Massachusetts General Hospital and is the recent recipient of the LBCA and Sarah Bieze ILC Research grant. Dr. Bahl will be working with ILC patient advocate, Beth Davis, through the yearlong study implementation period.
Dr. Bahl explains that her interest in ILC stems from her work on interval breast cancers. “Interval breast cancers” are breast cancers that are found and diagnosed sometime within the year following a negative screening mammogram and before a next screening mammogram when a patient presents with symptoms. She has found that lobular represents a disproportionate number of interval cancers. This drove her interest in finding ways to improve early detection, imaging techniques and more diagnostic accuracy for patients with ILC. Dr. Bahl believes that advancing research with this focus has the potential to enhance screening strategies, reduce misdiagnosis and ultimately improve patient outcomes.
Q&A with Dr. Manisha Bahl about her research study plan
Q. What is your patient population/where will you gather the screenings?
The study will train the artificial intelligence (AI) program using an algorithm developed from mammogram screening exams conducted in Mass General Hospital using tomosynthesis (which MGH began using in 2011). The AI program will be given the images of thousands of patients diagnosed with ILC initially and eventually and will be trained to pick up on all of the signs of ILC in the scans including those that were missed by the radiologists. The AI will be trained to always notice these signs or distortions signifying ILC in the images. MGH has a large database from exams going back over 10 years, 400,000 scans for training purposes. And we will validate, that is “test the accuracy of” the AI algorithm on another hospital’s imaging data (the Brigham) to confirm its accuracy.
Q. Do people have to authorize their scan being used?
No. We had to get approval to conduct the study which is using the imaging data collected in the past. It is not feasible to get written consent from every patient who had been seen but it also is not necessary or required because it is not impacting patient care and the images studied will have been “de-identified” before analysis, which means not connected to or connectable to specific patients’ names.
Q. If the study proves successful, will the AI program become the first or only thing analyzing a patient’s scan?
If the study proves to have provided the AI with an effective algorithm that has been validated with other hospitals’ imaging data, the AI will become a “decision support tool” for radiologists. All imaging research so far indicates that the most effective method for reviewing scans is achieved by machine and human working together. So the AI will be used as support, never on its own.
The vision of this study is that the algorithm will enable the AI to mark lesions on mammograms. There are then two ways it can be used. Radiologists can either choose to view what the AI said and then do their own analysis of the scan, or choose to minimize “automation bias” by looking at the scan first and THEN looking at the AI results. The automation bias occurs when the radiologist looks at the AI first and has begun to trust that the AI is always accurate and thus potentially scrutinize the scans less carefully. The AI will never be used alone.
Q. How long will this study be?
The grant study will run for one year and will start as soon as the database is ready to receive the scans.
Q. Will you present your findings at the San Antonio Breast Cancer Symposium 2025?
That is a possibility and so is the next ILC Symposium in 2026
Q. How many other people may be doing research on this?
There has definitely been interest in AI in radiology and there are algorithms that have been approved for clinical use, some on breast density, some for lesion detection for risk assessment during triage but this study will be the first that I am aware of training the AI to detect ILC.
Q. Are you using centralized pathology results in selecting which patients’ scans to use?
Yes. To create the algorithm, we want to be certain that we are using the scans from patients who have been diagnosed with ILC. The gold standard in determining the diagnosis of ILC is to use both the biopsy pathology and surgical pathology results. If for one patient there appears to be a difference in diagnosis between the biopsy and surgical results, we will use the surgical result diagnosis.
Q. Why did you decide to focus on this kind of imaging research?
I have found that ILC is a unique and fascinating subtype of breast cancer. All radiologists recognize how difficult ILC is to image (they refer to it as “the stealthy” breast cancer). My interest also comes from my desire to study Interval breast cancers and how to reduce them. An interval cancer is one that is found within a year of a “negative” screening mammogram. ILCs are disproportionately in this group because of the higher rate of false negative screens. I also feel that it is best to improve the efficacy of mammography so that people do not need to rely on an MRI.
Q. Will the algorithm train the AI to detect ILC better even in dense breasts?
AI does better than the human eye at finding breast cancers in dense breast tissue. There was a study that ran missed cancer cases through AI and they found that AI detected a higher amount of the false negatives in dense breast tissue. For some reason, it seems like AI is less affected than humans when viewing dense breast tissue
Q. Will you sort train the AI to know what to do depending on level of breast density?
Yes. We will train the AI on detection and can analyze how well it does focusing on many different factors including breast density, age, actual cancer types (hormone receptors), size, and more and determine how AI does at detecting the ILC in each scenario and for which it does better or worse.
Q. What are important ways that AI can be used to improve detection of ILC?
The AI training being developed will be used in conjunction with screening mammograms to better detect ILC reducing instances of when the screening would have missed the ILC tumor(s). Another way to train AI is to include mammograms that have positive ILC and then add in screenings for 10 years prior to train AI to see the change in breast overtime since ILC tends to grow for a while before being detected
Q. How do you go about preparing the files for training the AI/creating the algorithm?
Dr. Bahl’s research team collaborates with colleagues with AI expertise. There is established infrastructure through MGH to extract imaging information from the MGH databases. Dr. Bahl’s team, which includes a patient advocate to assist with the thinking through of their processes, submits a list of what mammograms they want and these are extracted from databases and sent to the lab
Q. If effective, how will this algorithm be shared with other hospitals?
AI is software, meaning the hospital IT team has to install it on computers. Hospitals are able to implement use of an AI algorithm they deem helpful. They don’t need FDA approval to use within the hospital system but a new screening algorithm or device must go through FDA approval process before it can be made available outside of the hospital system. The team will consider things like patents or seeking FDA approval down the road.
Q. Could you use another hospital’s screenings to train your AI?
Yes. They will do so in a way that enables MGH to ensure that no other hospital has free access to it once it has been proven effective for now. So they will use other MGH hospitals and other affiliated hospitals in different areas with different demographics to validate their algorithm such as affiliated hospital system (Brigham) which would have population differences.
