Avoiding Tissue Overlap in 2D Images: Single-Slice DBT Classification Using Convolutional Neural Networks

Abstract

Breast cancer was the most diagnosed cancer around the world in 2020. Screening programs, based on mammography, aim to achieve early diagnosis which is of extreme importance when it comes to cancer. There are several flaws associated with mammography, with one of the most important being tissue overlapping that can result in both lesion masking and fake-lesion appearance. To overcome this, digital breast tomosynthesis takes images (slices) at different angles that are later reconstructed into a 3D image. Having in mind that the slices are planar images where tissue overlapping does not occur, the goal of the work done here was to develop a deep learning model that could, based on the said slices, classify lesions as benign or malignant. The developed model was based on the work done by Muduli et. al, with a slight change in the fully connected layers and in the regularization done. In total, 77 DBT volumes-39 benign and 38 malignant-were available. From each volume, nine slices were taken, one where the lesion was most visible and four above/below. To increase the quantity and the variability of the data, common data augmentation techniques (rotation, translation, mirroring) were applied to the original images three times. Therefore, 2772 images were used for training. Data augmentation techniques were then applied two more times-one set used for validation and one set used for testing. Our model achieved, on the testing set, an accuracy of 93.2% while the values of sensitivity, specificity, precision, F1-score, and Cohen's kappa were 92%, 94%, 94%, 94%, and 0.86, respectively. Given these results, the work done here suggests that the use of single-slice DBT can compare to state-of-the-art studies and gives a hint that with more data, better augmentation techniques and the use of transfer learning might overcome the use of mammograms in this type of studies.

Keywords: CNN; DBT; carcinoma of the breast; deep learning.

Figure 1

Data selection and augmentation—image contrast modified just for illustration purposes. From each original volume, the most representative slice and the four slices above and below were extracted. The rotation was randomly preformed in a range from −50° to 50°, the translation could be horizontal or vertical, ranging from 1 pixel to 5% of the overall image size, and the reflection could either be vertical or horizontal.

Figure 2

Outline of the developed deep learning algorithm—composed by four convolution–batch normalization–ReLU–max pooling blocks, followed by two fully connected layers and a softmax layer.

Figure 3

Results of the image adjustment procedure. On the left, the original image is shown with a bounding box encapsulating the lesion. On the right, the same can be seen but for the image after adjusting. It can be perceived how the lesion becomes more visible.

Figure 4

ROI definition to encapsulate malignant (on the left) and benign (on the right) lesions.

Figure 5

Data augmentation procedure applied to the previously seen ROIs. Random transformations that include rotation and/or translation and/or mirroring were applied to each ROI. As it can be perceived, the images are substantially different from what can be observed in Figure 4, while still maintaining the lesion within the image limits.

Figure 6

Confusion matrix obtained with the validation set. From the total 693 instances, the model was capable of correctly classify 629 of them. It can also be noted that approximately 92% of the benign lesions were correctly classified, while for the malignant lesions, the same happened for approximately 89% of the instances.

Figure 7

Confusion matrix obtained with the test set. From the total 693 instances, the model was capable of correctly classify 646 of them. It can also be noted that approximately 92% of the benign lesions were correctly classified, while for the malignant lesions, the same happened for approximately 94% of them.