AI Improves Cancer Detection and Reading Time of Digital Breast Tomosynthesis

See also the article by Park et al in this issue.

Min Sun Bae, MD, PhD, is an associate professor in the breast imaging section in the department of radiology at Korea University Ansan Hospital. Her research interests include breast imaging, deep learning, and artificial intelligence. She is a board member of the Korean Society of Breast Imaging and serves on the editorial board of breast imaging for the Korean Journal of Radiology.

Digital breast tomosynthesis (DBT) has been widely accepted for breast cancer screening since its first approval by the U.S. Food and Drug Administration for clinical use in 2011 (1). DBT is an imaging technique that allows a three-dimensional (3D) reconstruction of the breast owing to the acquisition of multiple x-ray projection images at different angles. Mammography is the only screening modality that has been proven to reduce mortality from breast cancer by at least 20% (2). However, conventional two-dimensional (2D) mammography’s sensitivity and specificity are reduced by the presence of superimposed breast tissue, which may obscure or mimic malignant lesions. DBT reduces the masking effect of overlapping dense fibroglandular tissue, thereby improving cancer detection and reducing false-positive recalls. Studies have shown that screening with DBT substantially increases cancer detection rates and decreases recall rates compared with screening with digital mammography (DM) alone (3,4). Despite its many benefits, there are barriers to the wider use of DBT in clinical practice: increased equipment costs, larger image storage requirements, and longer interpretation time compared with DM (5).

The increased interpretation time is due to the added time for the radiologist to scroll through the reconstructed DBT images. Computer-aided detection (CAD) applications that flag lesion location in the DBT image stack could help in detecting suspicious findings while reducing interpretation time. When CAD was applied to DM, traditional CAD did not improve the sensitivity of radiologists and decreased their specificity because of a high false-positive rate. Recently, emerging CAD software based on artificial intelligence (AI) has been developed to improve mammography interpretation (6). A common focus for AI development for DM and DBT is breast cancer detection. Several studies have shown that the use of AI algorithms for DBT results in noninferior or improved sensitivity for cancer detection and reduced reading time (7–9).

In this issue of Radiology: Artificial Intelligence, Park et al (10) reaffirm the great impact of AI on diagnostic accuracy and reading time in the interpretation of DBT. The authors used a pretrained AI model trained on a large 2D mammogram dataset and fine-tuned the model on 12 810 DBT examinations from two vendors (Hologic, 10 436, 81.5%; GE HealthCare, 2374, 18.5%) using the ground truth labels of breast cancer. Their model was developed based on a deep convolutional neural network with a ResNet-34 backbone. The AI model provides four-view heatmaps for both synthetic 2D mammography and 3D sections of DBT and an abnormality score per lesion on each input image. Park and colleagues evaluated the performance of a stand-alone AI system and investigated the impact of AI support on the diagnostic accuracy and reading time of radiologists in DBT interpretation. The retrospective study included 2202 DBT examinations with a cancer prevalence of 50% (1102 cancer, 367 benign, and 733 normal) obtained from 14 imaging centers in the United States, including 258 cases for the reader study.

Park et al used a per-mammogram abnormality score, which was defined as the maximum score at four-view DBT to evaluate the stand-alone performance of an AI system. Additionally, the authors recruited 15 readers (eight general radiologists and seven breast imaging specialists), asking them to review DBT images from an enriched dataset of 258 cases that consisted of 65 (25%) cancers. Each reader determined whether to recall women from DBT. The readers localized the most suspicious lesion by drawing a contour of the lesion and provided suspicion scores for breast cancer. Reading time was defined as the time between when the readers clicked a start button to view images and when they selected a final decision. Outcome measures included the area under the receiver operating characteristic curve (AUC), sensitivity, specificity, and reading time.

The results of this study show that the use of AI for DBT interpretation can improve both accuracy and efficiency in breast cancer detection. The AUCs of a stand-alone AI system were 0.93 and 0.92 in the entire validation and reader study datasets, respectively. A remarkable finding was that the overall AUC of radiologists was significantly improved from 0.90 to 0.92 (P = .003) when interpreting DBT with AI assistance. Interestingly, the AUC of general radiologists improved from 0.89 to 0.92, which was comparable to the AUC of breast specialists with AI (0.92) and higher than that of breast specialists without AI (0.90). The overall sensitivity of radiologists was significantly higher with AI than without AI (87.7% vs 85.4%, P = .04), while no difference in the overall specificity was observed (79.6% vs 77.3%, P = .10). Park et al found that DBT reading time decreased for nine radiologists and increased for six radiologists. The mean reading time was 54.4 seconds for unaided readings and 48.5 seconds for AI-assisted readings (reading time difference, 5.9 seconds; P < .001). The authors concluded that an AI algorithm developed for detecting breast cancer at DBT significantly improved the diagnostic performance of radiologists, with the reading time being reduced.

Park et al (10) add to the existing evidence that the use of AI systems improves the performance of radiologists in interpreting DBT studies. In a previous study by Conant et al (8), 24 radiologists were asked to interpret all 260 cases with and without AI across two sessions. Conant et al found that the DBT AI system significantly reduced interpretation time by an average of 34.7 seconds (from 64.1 to 30.4 seconds), while improving sensitivity and specificity and reducing recall rates. Improving accuracy and time savings is of clear value in interpreting mammograms, especially in a screening setting. Mammogram interpretation continues to be one of the most difficult areas of radiology. A recent meta-analysis highlights the lack of studies on the role of AI in DBT (7). More studies regarding AI for DBT should be conducted in the future.

There were some limitations to the current study. Regarding the reader study, cancer-enriched datasets were retrospectively collected, and all 15 readers were from the United States. Therefore, the results reported by Park et al may not necessarily be applied to different regions or clinical settings. Further prospective studies are needed to determine whether the results of the AI system assessed are reproducible in different patient populations and screening circumstances. Another potential limitation was that the AI system was designed for DM and then fine-tuned using DBT examinations obtained from two vendors. This limits the generalizability of the AI system when attempting to apply it to multiple different vendors. At present, it is unclear how AI plays a role in breast imaging workflow and interpretation. The precise role of AI in real clinical practice is still to be determined.

In conclusion, this study has developed an AI algorithm using large-scale DBT examinations for breast cancer detection. The diagnostic accuracy of the stand-alone AI system was superior to that of radiologists. In a multireader multicase study, Park et al demonstrated that when AI is concurrently used in DBT interpretation, the reading times of radiologists can be reduced while diagnostic accuracy is improved. Given that DBT continues to replace 2D mammography, AI algorithms specific to DBT are necessary. The development and application of AI in breast imaging practice will continue to advance rapidly.