Artificial intelligence for the real world of breast screening

Performance of artificial intelligence in 7533 consecutive prevalent screening mammograms from the BreastScreen Australia program

OBJECTIVES

To assess the performance of an artificial intelligence (AI) algorithm in the Australian mammography screening program which routinely uses two independent readers with arbitration of discordant results.

METHODS

A total of 7533 prevalent round mammograms from 2017 were available for analysis. The AI program classified mammograms into deciles on the basis of breast cancer (BC) risk. BC diagnoses, including invasive BC (IBC) and ductal carcinoma in situ (DCIS), included those from the prevalent round, interval cancers, and cancers identified in the subsequent screening round two years later. Performance was assessed by sensitivity, specificity, positive and negative predictive values, and the proportion of women recalled by the radiologists and identified as higher risk by AI.

RESULTS

Radiologists identified 54 women with IBC and 13 with DCIS with a recall rate of 9.7%. In contrast, 51 of 54 of the IBCs and 12/13 cases of DCIS were within the higher AI score group (score 10), a recall equivalent of 10.6% (a difference of 0.9% (CI -0.03 to 1.89%, p = 0.06). When IBCs were identified in the 2017 round, interval cancers classified as false negatives or with minimal signs in 2017, and cancers from the 2019 round were combined, the radiologists identified 54/67 and 59/67 were in the highest risk AI category (sensitivity 80.6% and 88.06 % respectively, a difference that was not different statistically).

CONCLUSIONS

As the performance of AI was comparable to that of expert radiologists, future AI roles in screening could include replacing one reader and supporting arbitration, reducing workload and false positive results.

CLINICAL RELEVANCE STATEMENT

AI analysis of consecutive prevalent screening mammograms from the Australian BreastScreen program demonstrated the algorithm's ability to match the cancer detection of experienced radiologists, additionally identifying five interval cancers (false negatives), and the majority of the false positive recalls.

KEY POINTS

• The AI program was almost as sensitive as the radiologists in terms of identifying prevalent lesions (51/54 for invasive breast cancer, 63/67 when including ductal carcinoma in situ). • If selected interval cancers and cancers identified in the subsequent screening round were included, the AI program identified more cancers than the radiologists (59/67 compared with 54/67, sensitivity 88.06 % and 80.6% respectively p = 0.24). • The high negative predictive value of a score of 1-9 would indicate a role for AI as a triage tool to reduce the recall rate (specifically false positives).

Automated Triage of Screening Breast MRI Examinations in High-Risk Women Using an Ensemble Deep Learning Model

OBJECTIVES

The aim of the study is to develop and evaluate the performance of a deep learning (DL) model to triage breast magnetic resonance imaging (MRI) findings in high-risk patients without missing any cancers.

MATERIALS AND METHODS

In this retrospective study, 16,535 consecutive contrast-enhanced MRIs performed in 8354 women from January 2013 to January 2019 were collected. From 3 New York imaging sites, 14,768 MRIs were used for the training and validation data set, and 80 randomly selected MRIs were used for a reader study test data set. From 3 New Jersey imaging sites, 1687 MRIs (1441 screening MRIs and 246 MRIs performed in recently diagnosed breast cancer patients) were used for an external validation data set. The DL model was trained to classify maximum intensity projection images as "extremely low suspicion" or "possibly suspicious." Deep learning model evaluation (workload reduction, sensitivity, specificity) was performed on the external validation data set, using a histopathology reference standard. A reader study was performed to compare DL model performance to fellowship-trained breast imaging radiologists.

RESULTS

In the external validation data set, the DL model triaged 159/1441 of screening MRIs as "extremely low suspicion" without missing a single cancer, yielding a workload reduction of 11%, a specificity of 11.5%, and a sensitivity of 100%. The model correctly triaged 246/246 (100% sensitivity) of MRIs in recently diagnosed patients as "possibly suspicious." In the reader study, 2 readers classified MRIs with a specificity of 93.62% and 91.49%, respectively, and missed 0 and 1 cancer, respectively. On the other hand, the DL model classified MRIs with a specificity of 19.15% and missed 0 cancers, highlighting its potential use not as an independent reader but as a triage tool.

CONCLUSIONS

Our automated DL model triages a subset of screening breast MRIs as "extremely low suspicion" without misclassifying any cancer cases. This tool may be used to reduce workload in standalone mode, to shunt low suspicion cases to designated radiologists or to the end of the workday, or to serve as base model for other downstream AI tools.

Artificial Intelligence for Indication of Invasive Assessment of Calcifications in Mammography Screening

PURPOSE

Lesion-related evaluation of the diagnostic performance of an individual artificial intelligence (AI) system to assess mamographically detected and histologically proven calcifications.

MATERIALS AND METHODS

This retrospective study included 634 women of one screening unit (July 2012 - June 2018) who completed the invasive assessment of calcifications. For each leasion, the AI-system calculated a score between 0 and 98. Lesions scored > 0 were classified as AI-positive. The performance of the system was evaluated based on its positive predictive value of invasive assessment (PPV3), the false-negative rate and the true-negative rate.

RESULTS

The PPV3 increased across the categories (readers: 4a: 21.2 %, 4b: 57.7 %, 5: 100 %, overall 30.3 %; AI: 4a: 20.8 %, 4b: 57.8 %, 5: 100 %, overall: 30.7 %). The AI system yielded a false-negative rate of 7.2 % (95 %-CI: 4.3 %: 11.4 %) and a true-negative rate of 9.1 % (95 %-CI: 6.6 %; 11.9 %). These rates were highest in category 4a, 12.5 % and 10.4 % retrospectively. The lowest median AI score was observed for benign lesions (61, interquartile range (IQR): 45-74). Invasive cancers yielded the highest median AI score (81, IQR: 64-86). Median AI scores for ductal carcinoma in situ were: 74 (IQR: 63-84) for low grade, 70 (IQR: 52-79) for intermediate grade and 74 (IQR: 66-83) for high grade.

CONCLUSION

At the lowest threshold, the AI system yielded calcification-related PPV3 values that increased across categories, similar as seen in human evaluation. The strongest loss in AI-based breast cancer detection was observed for invasively assessed calcifications with the lowest suspicion of malignancy, yet with a comparable decrease in the false-positive rate. An AI-score based stratification of malignant lesions could not be determined.

KEY POINTS

· The AI-based PPV3 for calcifications is comparable to human assessment.. · AI showed a lower detection performance of screen-positive and screen-negative lesions in category 4a.. · Histological subgroups could not be discriminated by AI scores..

CITATION FORMAT

· Weigel S, Brehl AK, Heindel W et al. Artificial Intelligence for Indication of Invasive Assessment of Calcifications in Mammography Screening. Fortschr Röntgenstr 2023; 195: 38 - 46.

Comparison of Diagnostic Performance in Mammography Assessment: Radiologist with Reference to Clinical Information Versus Standalone Artificial Intelligence Detection

We compared diagnostic performances between radiologists with reference to clinical information and standalone artificial intelligence (AI) detection of breast cancer on digital mammography. This study included 392 women (average age: 57.3 ± 12.1 years, range: 30−94 years) diagnosed with malignancy between January 2010 and June 2021 who underwent digital mammography prior to biopsy. Two radiologists assessed mammographic findings based on clinical symptoms and prior mammography. All mammographies were analyzed via AI. Breast cancer detection performance was compared between radiologists and AI based on how the lesion location was concordant between each analysis method (radiologists or AI) and pathological results. Kappa coefficient was used to measure the concordance between radiologists or AI analysis and pathology results. Binominal logistic regression analysis was performed to identify factors influencing the concordance between radiologists’ analysis and pathology results. Overall, the concordance was higher in radiologists’ diagnosis than on AI analysis (kappa coefficient: 0.819 vs. 0.698). Impact of prior mammography (odds ratio (OR): 8.55, p < 0.001), clinical symptom (OR: 5.49, p < 0.001), and fatty breast density (OR: 5.18, p = 0.008) were important factors contributing to the concordance of lesion location between radiologists’ diagnosis and pathology results.