DBTex (Phase One) Submission
Faster R-CNN 2D (In-Slice) Detection Ensembling followed by 3D Reconstruction
Team Members
Akinyinka Omigbodun, Chrysostomos Marasinou, Noor Nakhaei, Anil Yadav, Nina Capiro, Bo Li, Anne Hoyt, William Hsu
Summary
The algorithm for our submission had three main stages:
- a first stage to detect biopsied tissue in DBT slices with a convolutional neural network (CNN) detector,
- followed by a stage to reconstruct biopsied tissue volumes from detections in slices, and
- a final stage to remove false-positive predictions with clinically significant skin markers using a detector with simple geometric rules.
Our algorithm achieved a mean sensitivity of 0.814 on the DBTex test set and placed 4th in the challenge.
2D (In-slice) Lesion Detection Ensembling
We utilized Faster R-CNN, a 2D detector with a Feature Pyramid Network (FPN) feature extractor and a Resnet-50 backbone. The model was initialized with weights from pre-training with the COCO train2017 data split. The input to the model was 3 consecutive slices that were concatenated as a 3-channel image. For training, slices around the lesion centers z (in the range, z-3 to z+3) were used with random flipping, brightness changes, and gamma adjustment augmentations applied. To monitor the performance during training, a portion of the training set was held out (20% patient-wise); The training set was split in 5, and 5 models were trained in a cross-validation scheme. The mean sensitivity on the held-out data was 0.780 ± 0.07 at 2 false positives per slice. During inference on the DBTex validation and testing data, all 5 models were applied to every slice, generating 5 sets of bounding box predictions for each slice. Detections from the 3 individually best of 5 models on the DBTex validation data were selected for the final ensembling [1], as the mean sensitivity was better than with all 5 models.
3D Lesion Candidate Generation
We then took the 2D in-slice bounding box predictions and combined them into 3D candidates based on:
- the proximity of slices in a DBT scan (being within 50% of the total number of slices of each other),
- 2D bounding box score threshold (at least 0.85 on a scale of 0 to 1), and
- 2D bounding boxes with an intersection over smaller intersecting box (IoSIB) greater than 0.75.
IoSIB gave better results than the standard intersection over union (IoU). 3D candidates with a depth of 1 (a single slice) were removed. The score assigned to a 3D candidate was the average of the top 10 scores of all 2D bounding boxes involved in its construction.
Blob Detection for False Positive Removal
To remove false positives (with clinical circular markers), we used the SimpleBlobDetector in the python OpenCV module (cv2) as summarized in the table below.
| Parameter | Value |
|---|---|
| filterbyArea | True |
| minArea | 1000 |
| filterByConvexity | True |
| minConvexity | 0.001 |
| maxConvexity | 0.4 |
| filterByInertia | True |
| minInertiaRatio | 0.01 |
| maxInertiaRatio | 1 |
| minThreshold | 50 ([0, 255] range) |
| maxThreshold | 150 ([0, 255] range) |
References
- Casado-García, Á., & Heras, J. (2020). Ensemble methods for object detection. In ECAI 2020 (pp. 2688-2695). IOS Press.