Automated and real-time segmentation of suspicious breast masses using convolutional neural network

Abstract

In this work, a computer-aided tool for detection was developed to segment breast masses from clinical ultrasound (US) scans. The underlying Multi U-net algorithm is based on convolutional neural networks. Under the Mayo Clinic Institutional Review Board protocol, a prospective study of the automatic segmentation of suspicious breast masses was performed. The cohort consisted of 258 female patients who were clinically identified with suspicious breast masses and underwent clinical US scan and breast biopsy. The computer-aided detection tool effectively segmented the breast masses, achieving a mean Dice coefficient of 0.82, a true positive fraction (TPF) of 0.84, and a false positive fraction (FPF) of 0.01. By avoiding positioning of an initial seed, the algorithm is able to segment images in real time (13-55 ms per image), and can have potential clinical applications. The algorithm is at par with a conventional seeded algorithm, which had a mean Dice coefficient of 0.84 and performs significantly better (P< 0.0001) than the original U-net algorithm.

Fig 1. Architecture of (a) Original U-net algorithm, (b) Multi U-net algorithm used to segment suspicious breast masses.

Multi U-net consists of 10 models architecturally similar to original U-net followed by majority voting to have a single binary segmentation mask.

Fig 2. Boxplot showing the performance of Multi U-net and DRLS algorithm for (a) Dice Coefficient, (b) TPF, and (c) FPF for benign, fibroadenoma, malignant, and invasive ductal carcinoma.

TPF indicates true positive fraction; FPF indicates false positive fraction.

Fig 3. Boxplot showing the performance of Multi U-net and DRLS algorithm for (a) Dice coefficient, (b) TPF, and (c) FPF for BI-RADS 3, 4 and 5. BI-RADS indicate Breast Imaging Reporting and Data System.

TPF indicates true positive fraction; FPF indicates false positive fraction.

Fig 4. Dice coefficient values for different majority voting thresholds along with error bars.

Fig 5

(a) B-mode Image of benign-cellular fibroepithelial Mass. (b) The manually segmented boundary is shown in red, the Multi U-net predicted boundary is shown in blue, the DRLS predicted boundary is shown in green and original U-net is shown in cyan.

Fig 6

(a) B-mode Image of benign fat necrosis. (b) Manually segmented boundary is shown in red, Multi U-net predicted boundary is shown in blue, the DRLS predicted boundary is shown in green and original U-net is shown in cyan.

Fig 7

(a) B-mode Image of invasive/infiltrating ductal carcinoma. (b) Manually segmented boundary is shown in red, Multi U-net predicted boundary is shown in blue, the DRLS predicted boundary is shown in green and original U-net is shown in cyan.

Fig 8

(a) B-mode Image of fibroadenoma with mild usual ductal hyperplasia and apocrine cysts. (b) Manually segmented boundary is shown in red, Multi U-net predicted boundary is shown in blue, the DRLS predicted boundary is shown in green and original U-net is shown in cyan.

Fig 9

(a) B-mode Image of invasive mammary carcinoma. (b) Manually segmented boundary is shown in red, Multi U-net predicted boundary is shown in blue, the DRLS predicted boundary is shown in green and original U-net is shown in cyan.

Fig 10

(a) B-mode Image of clustered apocrine cysts. (b) The manually segmented boundary is shown in red, Multi U-net predicted boundary is shown in blue, the DRLS predicted boundary is shown in green and original U-net is shown in cyan.

Fig 11

(a) Bar plot comparing Dice coefficient between original U-net and ten folds of Multi U-net. Each fold is evaluated 5 times to show the variance within each fold. (b) Error bars showing the increasing performance of U-net as more models are included in majority voting. Five different U-nets models are evaluated to show the variance.