Melissa L. Barron

Objectives: Breast cancer screening with mammography reduces mortality, but interpretation errors persist. While artificial intelligence (AI) shows promise, most studies focus on case-level accuracy and overlook lesion-level performance. This study evaluated the AI (GMIC + CLAHE) model against radiologists and radiology trainees, considering breast density and lesion characteristics. Methods: Nine BREAST test sets of screening digital mammograms (540 cases: 179 cancer, 361 normal) were analyzed. AI performance was compared with 17 radiologists and 26 trainees. Case- and lesion-level performance were assessed across breast density, lesion type, and size. AI malignancy scores were dichotomized using Youden’s index. Sensitivity, specificity, ROC AUC, lesion localization, and odds ratios (ORs) were calculated. Results: AI achieved higher case-level AUC (0.974) than radiologists (0.82 ± 0.06; P = .01) and trainees (0.74 ± 0.05; P < .0001) across breast densities. At lesion level, AI matched radiologists and outperformed trainees (OR = 1.6; P = .03), with lesion sensitivities of 66.3% (AI), 67.0% (radiologists), and 55.0% (trainees). AI also outperformed trainees in detecting mixed-type lesions (75% vs 55%; OR = 2.4) and small lesions ≤15 mm (65.7% vs 55%; OR = 1.6). Architectural distortion had the highest miss rate (43.8%) whereas mixed-type lesions demonstrated the lowest miss rate (25.0%) by AI. Conclusions: AI demonstrated superior case-level accuracy compared with radiologists and trainees, but limitations in localizing specific lesion types, supporting its role as an adjunct to human readers in screening and training. Advances in knowledge: This study integrates case-and-lesion-level evaluation in different case characteristics, identifying key strengths and limitations for clinical implementation. Despite superior case-level accuracy, variability in lesion localization across cancer types and sizes highlights important considerations for its integration into screening services.

Background: This study investigated the impact of Digital Breast Tomosynthesis (DBT) training workshops on doctors' performance in simultaneous‐double‐reader scenarios. Methods: Ten pairs of Vietnamese readers, including radiologists, registrars, and breast physicians, participated in the workshop, which featured lectures and breast image reading sessions provided by Australian experts. The first session included a test set of 30 screening full‐field digital mammograms (FFDM) (10 cancers and 20 normal cases) with DBT images provided during the answer review stage. The second session with 35 cases (15 cancer) and session 3 with 30 cases (11 cancer) consisted of screening FFDM mammograms, DBT slices, and synthesized images. Participants used the BREAST‐VIETRAD platform to view breast images and detect cancer lesions using the RANZCR‐BIRADS scale. The study assessed performance disparities between the first and second DBT sets using the Wilcoxon Signed Rank test and explored the correlation between score changes and reader experience. Results: There were significant improvements in the readers' sensitivity (0.553 vs. 0.91; p = 0.005) and lesion sensitivity (0.419 vs. 0.709; p = 0.005) from the first to the second DBT set, matching the sensitivity seen in FFDM sets. This improvement was consistent across both low and high breast density cases. Notable enhancements in lesion sensitivity were also observed for detecting masses (0.340 vs. 0.633; p = 0.011), calcifications, and architectural distortions (0.450 vs. 0.933; p = 0.011). The probability of score improvement (ROC AUC and JAFROC FOM) in DBT of pairs of readers with both less than 3 years of experience in reading mammograms is five times greater than those with over 5 years of experience. Conclusions: Training sessions with simultaneous‐double readers significantly improved Vietnamese clinicians' breast cancer detection capabilities, highlighting the importance of tailored educational programs to enhance diagnostic accuracy in regions lacking formal screening initiatives.

Objectives: This study explored the familiarity, perceptions and confidence of Australian radiology clinicians involved in reading screening mammograms, regarding artificial intelligence (AI) applications in breast cancer detection. Methods: Sixty-five radiologists, breast physicians and radiology trainees participated in an online survey that consisted of 23 multiple choice questions asking about their experience and familiarity with AI products. Furthermore, the survey asked about their confidence in using AI outputs and their preference for AI modes applied in a breast screening context. Participants’ responses to questions were compared using Pearson’s ?2 test. Bonferroni-adjusted significance tests were used for pairwise comparisons. Results: Fifty-five percent of respondents had experience with AI in their workplaces, with automatic density measurement powered by machine learning being the most familiar AI product (69.4%). The top AI outputs with the highest ranks of perceived confidence were ‘Displaying suspicious areas on mammograms with the percentage of cancer possibility’ (67.8%) and ‘Automatic mammogram classification (normal, benign, cancer, uncertain)’ (64.6%). Radiology and breast physicians preferred using AI as second-reader mode (75.4% saying ‘somewhat happy’ to ‘extremely happy’) over triage (47.7%), pre-screening and first-reader modes (both with 26.2%) (P < 0.001). Conclusion: The majority of screen readers expressed increased confidence in utilising AI for highlighting suspicious areas on mammograms and for automatically classifying mammograms. They considered AI as an optimal second-reader mode being the most ideal use in a screening program. The findings provide valuable insights into the familiarities and expectations of radiologists and breast clinicians for the AI products that can enhance the effectiveness of the breast cancer screening programs, benefitting both healthcare professionals and patients alike.

This paper investigates the adaptability of four state-of-the-art artificial intelligence (AI) models to the Australian mammographic context through transfer learning, explores the impact of image enhancement on model performance and analyses the relationship between AI outputs and histopathological features for clinical relevance and accuracy assessment. A total of 1712 screening mammograms (n = 856 cancer cases and n = 856 matched normal cases) were used in this study. The 856 cases with cancer lesions were annotated by two expert radiologists and the level of concordance between their annotations was used to establish two sets: a ‘high-concordances subset’ with 99% agreement of cancer location and an ‘entire dataset’ with all cases included. The area under the receiver operating characteristic curve (AUC) was used to evaluate the performance of Globally aware Multiple Instance Classifier (GMIC), Global-Local Activation Maps (GLAM), I&H and End2End AI models, both in the pretrained and transfer learning modes, with and without applying the Contrast Limited Adaptive Histogram Equalization (CLAHE) algorithm. The four AI models with and without transfer learning in the high-concordance subset outperformed those in the entire dataset. Applying the CLAHE algorithm to mammograms improved the performance of the AI models. In the high-concordance subset with the transfer learning and CLAHE algorithm applied, the AUC of the GMIC model was highest (0.912), followed by the GLAM model (0.909), I&H (0.893) and End2End (0.875). There were significant differences (p < 0.05) in the performances of the four AI models between the high-concordance subset and the entire dataset. The AI models demonstrated significant differences in malignancy probability concerning different tumour size categories in mammograms. The performance of AI models was affected by several factors such as concordance classification, image enhancement and transfer learning. Mammograms with a strong concordance with radiologists’ annotations, applying image enhancement and transfer learning could enhance the accuracy of AI models.

This paper investigates the adaptability of four state-of-the-art Artificial Intelligence (AI) models to the Australian mammographic context through transfer learning, explores the impact of image enhancement on model performance and analyses the relationship between AI outputs and histopathological features for clinical relevance and accuracy assessment. A total of 1712 screening mammograms (n=856 cancer cases and n=856 matched normal cases) were used in this study. The 856 cases with cancer lesions were annotated by two expert radiologists and the level of concordance between their annotations was used to establish image subsets. The area under the receiver operating characteristic curve (AUC) was used to evaluate the performance of Globally aware Multiple Instance Classifier (GMIC), Global-Local Activation Maps (GLAM), I&H and End2End AI models, both in the pre-trained and transfer learning modes, with and without applying the Contrast Limited Adaptive Histogram Equalization (CLAHE) algorithm. The four AI models with and without transfer learning in the high-concordance subset outperformed those in the entire dataset. Applying the CLAHE algorithm to mammograms improved the performance of the AI models. In the high-concordance subset with transfer learning and CLAHE algorithm applied, the AUC of the GMIC model was highest (0.912), followed by GLAM model (0.909), I&H (0.893) and End2End (0.875). There were significant differences (P<0.05) in the performances of the four AI models between high-concordance subset and entire dataset. The AI models demonstrated significant differences in malignancy probability concerning different tumour size categories in mammograms. The performance of AI models was affected by several factors such as concordance classification, image enhancement and transfer learning. Mammograms with strong concordance of radiologists’ annotations, applying image enhancement and transfer learning could enhance the accuracy of AI models.

Objective: The study aims to evaluate the diagnostic efficacy of radiologists and radiology trainees in digital breast tomosynthesis (DBT) alone vs DBT plus synthesized view (SV) for an understanding of the adequacy of DBT images to identify cancer lesions. Methods: Fifty-five observers (30 radiologists and 25 radiology trainees) participated in reading a set of 35 cases (15 cancer) with 28 readers reading DBT and 27 readers reading DBT plus SV. Two groups of readers had similar experience in interpreting mammograms. The performances of participants in each reading mode were compared with the ground truth and calculated in term of specificity, sensitivity, and ROC AUC. The cancer detection rate in various levels of breast density, lesion types and lesion sizes between ‘DBT’ and ‘DBT + SV’ were also analyzed. The difference in diagnostic accuracy of readers between two reading modes was assessed using Man–Whitney U test. p < 0.05 indicated a significant result. Results: There was no significant difference in specificity (0.67-vs-0.65; p = 0.69), sensitivity (0.77-vs-0.71; p = 0.09), ROC AUC (0.77-vs-0.73; p = 0.19) of radiologists reading DBT plus SV compared with radiologists reading DBT. Similar result was found in radiology trainees with no significant difference in specificity (0.70-vs-0.63; p = 0.29), sensitivity (0.44-vs-0.55; p = 0.19), ROC AUC (0.59-vs-0.62; p = 0.60) between two reading modes. Radiologists and trainees obtained similar results in two reading modes for cancer detection rate with different levels of breast density, cancer types and sizes of lesions (p > 0.05). Conclusion: Findings show that the diagnostic performances of radiologists and radiology trainees in DBT alone and DBT plus SV were equivalent in identifying cancer and normal cases. Advances in knowledge: DBT alone had equivalent diagnostic accuracy as DBT plus SV which could imply the consideration of using DBT as a sole modality without SV.