Purpose: Screening of dense breasts has become a major concern as accuracy of mammography in dense tissue is low and results in inconclusive results. Due to this, many states in USA have introduced density reporting law requiring mammography health care providers to include appropriate information about breast density in the report. This makes more than 30% women have an inconclusive mammogram, despite going through the X-Ray based test. In this work, we evaluated a new method to automatically detect dense breasts without the need for patient to go through mammography. This would help in identifying women who need an alternate test for breast screening without putting them through a radiation-based test, also saving them from the anxiousness of an inconclusive result. Methods: We have developed a new AI-based technique to analyse different views of thermal images of patients and identify dense breasts. The heat generated from the chest wall reaches the breast surface by transmitting through different layers of breast tissue. A fatty tissue absorbs more heat due to their low conductivity whereas a dense tissue conducts more heat. These variations in conductivity result in variation of heat on the surface of the skin, depending on the ratio of fatty to dense tissues in the breast. Infrared cameras today can detect temperature variations of up to 0.05 deg C. Hence, analysis of multiple views of a patient thermal images can give a good signal about density of breast.
A retrospective analysis of 408 subjects with multiview thermal images was performed to evaluate the efficacy of the AI algorithm. All these patients are walk-ins into reputed cancer hospitals, who came either for regular screening or with symptoms. Women with prior cancer history, pregnancy and lactating are not included in this study. The true density classification of these subjects were captured through assessment of their mammograms and compared with the output of the algorithm. Out of these 408 subjects, 204 subjects were interpreted by radiologists to have dense breasts with BIRADS categories C and D and remaining 204 subjects were interpreted as fatty breasts with BIRADS categories A and B. Every subject underwent thermal imaging before mammography. Thermal images at 3 different view angles of 0, 45 and -45 deg were captured per patient.
A deep learning architecture using ResNet 50 was trained with these breast thermal images to predict the breast density per view. The Resnet probabilities of all the views for a subject were combined using a support vector machine (SVM) classifier to get a final prediction of the density for each subject. To improve the overall performance, age and some histogram-based features were also used along with Resnet probabilities to SVM classifier. Results: A five-fold cross validation, where the entire data was divided into 5 folds with 4 folds used to train the model and the remaining fold to test the model is performed to evaluate the results. We obtained an average area under curve (AUC) of 0.76 after repeating the experimentation with each of five-fold as test dataset. The algorithm achieved an average specificity of 88%, sensitivity of 50%, PPV of 81% and NPV of 64%. Conclusion: This is a preliminary study showing promising results in detecting non-dense breasts with 88% accuracy. This test can be performed before mammography and those found positive for density can potentially be referred to non-mammo test. This will help limit the number of women going for radiation-based screening when their result is likely to be birad0. Though the sensitivity is low, since these subjects would be going for mammogram anyhow, the person executes normal screening procedure. This could also reduce the cost and panic that can happen with inconclusive mammography reports.
Citation Format: Divyam Srivastava, Atishay Ganesh, Siva Teja Kakileti, Geetha Manjunath. A non-radiation based screening to detect dense breasts [abstract]. In: Proceedings of the 2019 San Antonio Breast Cancer Symposium; 2019 Dec 10-14; San Antonio, TX. Philadelphia (PA): AACR; Cancer Res 2020;80(4 Suppl):Abstract nr P2-03-06.