Developing breast lesion detection algorithms for digital breast tomosynthesis: Leveraging false positive findings

Abstract

Background: Due to the complex nature of digital breast tomosynthesis (DBT) in imaging techniques, reading times are longer than 2D mammograms. A robust computer-aided diagnosis system in DBT could help radiologists reduce their workload and reading times.

Purpose: The purpose of this study was to develop algorithms for detecting biopsy-proven breast lesions on DBT using multi-depth level convolutional models and leveraging non-biopsied samples. As biopsied positive samples in a lesion dataset are limited, we hypothesized that false positive (FP) findings by detection algorithms from non-biopsied benign lesions could improve detection algorithms by using them as data augmentation.

Approach: We first extracted 2D slices from DBT volumes with biopsy-proven breast lesions (cancer and benign), with non-biopsied benign lesions (actionable), and for controls. Then, to provide lesion continuity along the z-direction, we combined a lesion slice with its immediate adjacent slices to synthesize 2.5-dimensional (2.5D) images of the lesion by assigning them into R, G, and B color channels. We used 224 biopsy-proven lesions from 39 cancer and 62 benign patients from a DBTex challenge dataset of 1000 scans. We included the 2.5D images of immediate neighboring slices from the lesion's center to increase the number of training samples. For lesion detection, we used the YOLOv5 algorithm as our base network. We trained a baseline algorithm (medium-depth level) using biopsied samples to detect actionable FPs in non-biopsied images. Afterward, we fine-tuned the baseline model on the augmented image set (actionable FPs added). For lesion inferencing, we processed the DBT volume slice-by-slice to estimate bounding boxes in each slice, and then combined them by connecting bounding boxes along the depth via volumetric morphological closing. We trained an additional model (large) with deeper-depth levels by repeating the above process. Finally, we developed an ensemble algorithm by combining the medium and large detection models. We used the free-response operating characteristic curve to evaluate our algorithms. We reported mean sensitivity per FPs per DBT volume only for biopsied views and sensitivity at 2-false positives per image (2FPI) for all views. However, due to the limited accessibility to the truth of the challenge validation and test datasets, we used sensitivity at 2FPI for statistical evaluation.

Results: For the DBTex independent validation set, the medium baseline model achieved a mean sensitivity of 0.627 FPs per DBT volume, and a sensitivity of 0.640 at 2FPI. After adding actionable FP lesions, the model had an improved 2FPI of 0.769 over the baseline (p-value = 0.013). Our ensemble algorithm with multi-depth levels (medium + large) achieved a mean sensitivity of 0.815 FPs per DBT volume and an improved sensitivity at 2FPI of 0.80 over the baseline (p-value < 0.001) on the validation set. Finally, our ensemble model achieved a mean sensitivity of 0.786 FPs per DBT volume and a sensitivity of 0.743 at 2FPI on the DBTex independent test set.

Conclusions: Our results show that actionable FP findings hold useful information for lesion detection algorithms, and our ensemble detection model with multi-depth levels improves lesion detection performance.

Keywords: breast cancer; computer-aided detection; deep learning; digital breast tomosynthesis; lesion detection.

Figure 1.

Steps for preprocessing DBT voulmes. First row shows the procedure of the breast region segmentaion (slice-by-slice). Second row shows steps for 2.5D DBT image synthesis, combining the center lesion slice (n) with its neighboring lesions into three channels of RGB image format, and then all slices were resized to 640×640 to fit for CNN detector.

Figure 2.

Detecting FP lesions from non-biopsied images. The baseline DBT model takes non-biospied images as input and produces FP lesions with class label, score and bounding box coordinate (A). Some examples of detected FP lesions in actionable images for LCC and LMLO views (B).

Figure 3.

Conceptual diagram of our framework for breast lesion detection on DBT using false positive (FP) findings. We augmented the biopsied samples (2.5D images in our augmented DBTex training set) of the original DBTex by 3 times using adjacent slices of the center lesion following our image preprocessing method. The baseline models with different depth levels were developed using biopsied samples (cancer and benign) as positive samples. Here, s, m, l and x are YOLOv5 algorithms with different depth levels (Model_S, Model_M, Model_L and Model_X, respectively) by changing the number of convolution layers. One of the baseline models was used to detect false postive (FP) findings from non-biopsed samples. Then the baseline models were fine-tuned using augmented training images (FP findings as negative samples added with the biopseid samples). The internal-validation subset from our training set was used to optimize hyper parameters for both the baseline models development and during their fine-tuning. Finally, an ensemble algorihtm was developed with a combination of the above multi-depth levels models.

Figure 4.

Method of 2.5D lesion detection. We performed volumetric morphological closing using a 5×5×5 cube stucture. For lesion localization, the membership was consturcted using neighboring voxels (bwlabel function, MATLAB) in binary images. For all slices with membership I, we selected the maximum slice score as final score of a membership I (right).