Deep learning of longitudinal mammogram examinations for breast cancer risk prediction

Dadsetan, Saba; Arefan, Dooman; Berg, Wendie A.; Zuley, Margarita L.; Sumkin, Jules H.; Wu, Shandong

doi:10.1016/j.patcog.2022.108919

Cited by 16 publications

(9 citation statements)

References 35 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…Mammogram images are generally large in size, but due to limitations in GPU resources, it is often necessary to downsample the images to a smaller size for general deep learning models. Mammogram image sizes used in previous mammography-based risk prediction studies varied from 224×224 20,44 , 256×256 45 , 299×299 17 , 632×512 46 , and 2048×1664 19,21,23 . Considering the trade-off of the GPU memory and the number of historical mammograms used in this study, the input mammograms are resized to the size of 1024×512.…”

Section: Discussionmentioning

confidence: 99%

“…Performances in clinical practice remain modest, as they lack sensitivity to short/middle-term cancer risk variation due to the limited incorporation of individual-specific risk factors (detailed imaging characteristics) beyond breast density. In recent years, deep learning (DL) methods combining large screening mammography datasets with detailed risk factors have shown promise in balancing the harmto-benefit ratios of BC screening 7,[16][17][18][19][20][21][22] and have even been validated in clinical settings 23 . For instance, the recent MIRAI risk model 19 achieved state-of-the-art (SOTA) performance in five-year BC risk prediction and outperformed the clinically adopted traditional models 19,[23][24][25] .…”

Section: Introductionmentioning

confidence: 99%

“…Therefore, beyond learning risk features (e.g. breast density) from single-time point imaging, multi-time point learning may also be helpful in discovering the underlying dynamics of the risk pattern for BC development 20,27 . Besides, according to current BC guidelines, it is recommended that women who have undergone breast-conserving therapy for early-stage BC should continue with screening mammography 28,29 .…”

Section: Introductionmentioning

confidence: 99%

See 2 more Smart Citations

Predicting long term breast cancer risk using longitudinal mammographic screening history

Tan,

Wang,

Gao

et al. 2023

Preprint

View full text Add to dashboard Cite

Risk assessment of breast cancer (BC) seeks to enhance individualized screening and prevention strategies. Recent deep learning (DL) risk models based on mammography have shown superiority in short-term risk prediction compared to traditional risk factor-based models. However, these models primarily rely on single-time exams, which emphasize the detection of existing lesions and may ignore the temporal changes in breast tissues. In this study, we present the Multi-Time Point Breast Cancer Risk Model (MTP-BCR), a novel temporospatial DL risk model that integrates traditional BC risk factors and longitudinal mammography data to identify subtle changes in breast tissue indicative of future malignancy. Utilizing a large inhouse dataset comprising risk factors and 171,168 mammograms involving 9,133 women, we evaluate the performance of the MTP-BCR model in long-term risk prediction. Our model demonstrates a significant improvement in 10-year risk prediction with an area under the receiver operating characteristics (AUC) of 0.80, outperforming the traditional BCSC 10-year risk model, our pure image model (without risk factors), and is also superior to other SOTA methods at 5-year AUC in various screening cohorts. Furthermore, MTP-BCR provides unilateral breast-level predictions with AUCs up to 0.81 and 0.77 for 5-year and 10-year risk assessments, respectively. External validation in the public CSAW-CC dataset demonstrates the consistent advantage of our multi-time point-based model compared to the single-time point-based method. For the prediction of breast cancer recurrence, our MTP-BCR obtains a 5-year AUC of 0.71 which also surpasses other methods. The heatmaps derived from our model may help clinicians better understand the progression from normal tissue to cancerous growth, enhancing interpretability in breast cancer risk assessment.

show abstract

Section: Discussionmentioning

confidence: 99%

Section: Introductionmentioning

confidence: 99%

Section: Introductionmentioning

confidence: 99%

See 1 more Smart Citation

Predicting long term breast cancer risk using longitudinal mammographic screening history

Tan,

Wang,

Gao

et al. 2023

Preprint

View full text Add to dashboard Cite

show abstract

“…Only a few models overlook the potential value of considering longitudinal mammographic changes 3 where both historical and current mammograms are utilized as inputs in their deep learning model. In previous studies 4 , 5 by incorporating longitudinal mammogram examinations, the specifically designed deep learning models have shown better performance for breast cancer risk predicting task, which shows that changes in the breast tissue over time can provide critical insights into the risk of breast cancer.…”

Section: Introductionmentioning

confidence: 99%

Incorporating longitudinal changes of mammograms for breast cancer diagnosis

Zhou,

Arefan,

Zuley

et al. 2024

Medical Imaging 2024: Computer-Aided Diagnosis

View full text Add to dashboard Cite

In recent years, deep learning has showcased substantial promise in breast cancer diagnosis via mammograms. However, the integration of longitudinal changes between consecutive mammograms, which clinicians frequently consider for diagnosis, remains under-explored. In this study, we introduce an novel method that leverages crossattention mechanisms to capture longitudinal information between consecutive mammograms taken at various time intervals. Our method's efficacy was assessed using a case-control internal dataset consisting of 590 cases. Preliminary results underscore its superiority over models relying solely on a single "current" mammogram exam and those that merely combine features extracted from "current" and "prior" mammograms. By harnessing the power of longitudinal data, our model achieved enhanced diagnostic performance.

show abstract

“…The performances of these risk models remain modest in clinical practice and are not very sensitive to short/middle-term cancer risk variation due to the shortage of individual-specific risk adaptation, for example through the incorporation of detailed imaging findings beyond breast density only. With the recent boost in deep learning (DL) methods, some studies that combined large screening mammography datasets with detailed risk factors have shown considerable promise to help balance the harm-to-benefit ratios of BC screening (7,(16)(17)(18)(19)(20)(21)(22) and were even validated in clinical settings (23). For example, a recent study (19) developed a risk model, MIRAI, that achieved state-of-the-art (SOTA) performance in five-year BC risk prediction and outperformed the clinically adopted traditional models (19,(23)(24)(25).…”

Section: Introductionmentioning

confidence: 99%

Predicting up to 10 year breast cancer risk using longitudinal mammographic screening history

Wang

Tan

Gao

et al. 2023

Preprint

View full text Add to dashboard Cite

Risk assessment of breast cancer (BC) seeks to enhance individualized screening and prevention strategies. BC risk informs healthy individuals of the short- and long-term likelihood of cancer development, also enabling detection of existing BC. Recent mammographic-based deep learning (DL) risk models outperform traditional risk factor-based models and achieve state-of-the-art (SOTA) at short-term risk prediction, but mainly use single-time exams, which seem to rely more on detecting existing lesions. We present a novel temporospatial and explainable deep learning risk model, the Multi-Time Point Breast Cancer Risk Model (MTP-BCR), which learns from longitudinal mammography data to identify subtle changes in breast tissue that may signal future malignancy. Utilizing a large in-house dataset of 171,168 screening mammograms from 42,792 consecutive exams involving 9,133 women, our model demonstrates a significant improvement in long-term (10-year) risk prediction with an area under the receiver operating characteristics (AUC) of 0.80, outperforming the traditional BCSC 10-year risk model and other SOTA methods at 5-year AUC in various screening cohorts. Furthermore, MTP-BCR provides unilateral breast-level predictions, achieving AUCs up to 0.81 and 0.77 for 5-year risk and 10-year risk assessments, respectively. The heatmaps derived from our model may help clinicians better understand the progression from normal tissue to cancerous growth, enhancing interpretability in breast cancer risk assessment.

show abstract

Deep learning of longitudinal mammogram examinations for breast cancer risk prediction

Cited by 16 publications

References 35 publications

Predicting long term breast cancer risk using longitudinal mammographic screening history

Predicting long term breast cancer risk using longitudinal mammographic screening history

Incorporating longitudinal changes of mammograms for breast cancer diagnosis

Predicting up to 10 year breast cancer risk using longitudinal mammographic screening history

Contact Info

Product

Resources

About