Reliability of Computer-Assisted Breast Density Estimation: Comparison of Interactive Thresholding, Semiautomated, and Fully Automated Methods

Comparison of breast density assessment between human eye and automated software on digital and synthetic mammography: Impact on breast cancer risk

PURPOSE

To evaluate the agreement between automatic assessment software of breast density based on artificial intelligence (AI) and visual assessment by a senior and a junior radiologist, as well as the impact on the assessment of breast cancer risk (BCR) at 5 years.

MATERIALS AND METHODS

We retrospectively included 311 consecutive women (mean age, 55.6±8.5 [SD]; range: 40-74 years) without a personal history of breast cancer who underwent routine mammography between January 1, 2019 and February 28, 2019. Mammographic breast density (MBD) was independently evaluated by a junior and a senior reader on digital mammography (DM) and synthetic mammography (SM) using BI-RADS (5th edition) and by an AI software. For each MBD, BCR at 5 years was estimated per woman by the AI software. Interobserver agreement for MBD between the two readers and the AI software were evaluated by quadratic κ coefficients. Reproducibility of BCR was assessed by intraclass correlation coefficient (ICC).

RESULTS

Agreement for MBD assessment on DM and SM was almost perfect between senior and junior radiologists (κ=0.88 [95% CI: 0.84-0.92] and κ=0.86 [95% CI: 0.82-0.90], respectively) and substantial between the senior radiologist and AI (κ=0.79; 95% CI: 0.73-0.84). There was substantial agreement between DM and SM for the senior radiologist (κ=0.79; 95% CI: 0.74-0.84). BCR evaluation at 5 years was highly reproducible between the two radiologists on DM and SM (ICC=0.98 [95% CI: 0.97-0.98] for both), between BCR evaluation based on DM and SM evaluated by the senior (ICC=0.96; 95% CI: 0.95-0.97) or junior radiologist (ICC=0.97; 95% CI: 0.96-0.98) and between the senior radiologist and AI (ICC=0.96; 95% CI: 0.95-0.97).

CONCLUSION

This preliminary study demonstrates a very good agreement for BCR evaluation based on the evaluation of MBD by a senior radiologist, junior radiologist and AI software.

Determination of mammographic breast density using a deep convolutional neural network

OBJECTIVE

High breast density is a risk factor for breast cancer. The aim of this study was to develop a deep convolutional neural network (dCNN) for the automatic classification of breast density based on the mammographic appearance of the tissue according to the American College of Radiology Breast Imaging Reporting and Data System (ACR BI-RADS) Atlas.

METHODS

In this study, 20,578 mammography single views from 5221 different patients (58.3 ± 11.5 years) were downloaded from the picture archiving and communications system of our institution and automatically sorted according to the ACR density (a-d) provided by the corresponding radiological reports. A dCNN with 11 convolutional layers and 3 fully connected layers was trained and validated on an augmented dataset. The model was finally tested on two different datasets against: i) the radiological reports and ii) the consensus decision of two human readers. None of the test datasets was part of the dataset used for the training and validation of the algorithm.

RESULTS

The optimal number of epochs was 91 for medio-lateral oblique (MLO) projections and 94 for cranio-caudal projections (CC), respectively. Accuracy for MLO projections obtained on the validation dataset was 90.9% (CC: 90.1%). Tested on the first test dataset of mammographies (850 MLO and 880 CC), the algorithm showed an accordance with the corresponding radiological reports of 71.7% for MLO and of 71.0% for CC. The agreement with the radiological reports improved in the differentiation between dense and fatty breast for both projections (MLO = 88.6% and CC = 89.9%). In the second test dataset of 200 mammographies, a good accordance was found between the consensus decision of the two readers on both, the MLO-model (92.2%) and the right craniocaudal-model (87.4%). In the differentiation between fatty (ACR A/B) and dense breasts (ACR C/D), the agreement reached 99% for the MLO and 96% for the CC projections, respectively.

CONCLUSIONS

The dCNN allows for accurate classification of breast density based on the ACR BI-RADS system. The proposed technique may allow accurate, standardized, and observer independent breast density evaluation of mammographies.

ADVANCES IN KNOWLEDGE

Standardized classification of mammographies by a dCNN could lead to a reduction of falsely classified breast densities, thereby allowing for a more accurate breast cancer risk assessment for the individual patient and a more reliable decision, whether additional ultrasound is recommended.

High throughput nonparametric probability density estimation

In high throughput applications, such as those found in bioinformatics and finance, it is important to determine accurate probability distribution functions despite only minimal information about data characteristics, and without using human subjectivity. Such an automated process for univariate data is implemented to achieve this goal by merging the maximum entropy method with single order statistics and maximum likelihood. The only required properties of the random variables are that they are continuous and that they are, or can be approximated as, independent and identically distributed. A quasi-log-likelihood function based on single order statistics for sampled uniform random data is used to empirically construct a sample size invariant universal scoring function. Then a probability density estimate is determined by iteratively improving trial cumulative distribution functions, where better estimates are quantified by the scoring function that identifies atypical fluctuations. This criterion resists under and over fitting data as an alternative to employing the Bayesian or Akaike information criterion. Multiple estimates for the probability density reflect uncertainties due to statistical fluctuations in random samples. Scaled quantile residual plots are also introduced as an effective diagnostic to visualize the quality of the estimated probability densities. Benchmark tests show that estimates for the probability density function (PDF) converge to the true PDF as sample size increases on particularly difficult test probability densities that include cases with discontinuities, multi-resolution scales, heavy tails, and singularities. These results indicate the method has general applicability for high throughput statistical inference.