A Comparative Study of Automated Deep Learning Segmentation Models for Prostate MRI

Abstract

Prostate cancer is one of the most common forms of cancer globally, affecting roughly one in every eight men according to the American Cancer Society. Although the survival rate for prostate cancer is significantly high given the very high incidence rate, there is an urgent need to improve and develop new clinical aid systems to help detect and treat prostate cancer in a timely manner. In this retrospective study, our contributions are twofold: First, we perform a comparative unified study of different commonly used segmentation models for prostate gland and zone (peripheral and transition) segmentation. Second, we present and evaluate an additional research question regarding the effectiveness of using an object detector as a pre-processing step to aid in the segmentation process. We perform a thorough evaluation of the deep learning models on two public datasets, where one is used for cross-validation and the other as an external test set. Overall, the results reveal that the choice of model is relatively inconsequential, as the majority produce non-significantly different scores, apart from nnU-Net which consistently outperforms others, and that the models trained on data cropped by the object detector often generalize better, despite performing worse during cross-validation.

Keywords: deep learning; prostate cancer; prostate detection; prostate segmentation.

Figure A1

Evolution of the different metrics during the training of the prostate detection model. The figure contains the Recall, Precision, mAP@$0.5$, mAP@$0.5:0.95$, and both the object and box loss for the training and validation sets.

Figure 1

Comparison between original and cropped volumes. Both images correspond to the middle slice of the volumes. The image on the left represents the original full size slice, while the image on the right represents the cropped version.

Figure 2

Example of a batch of 16 images, and their respective segmentation masks, from the ProstateX dataset after being pre-processed and augmented. Each image corresponds to the tenth slice of the volume. The images show the effect of the affine transformations as well as the random bias field augmentations.

Figure 3

Comparison between original prostate images, masks and predicted volumes. The top row consists of random slices from different volumes, the second row consists of the respective prostate masks, and the third row consists of the respective predicted bounding boxes, with confidence value of the prostate on each of the images.

Figure 4

Boxplots showing the distribution of Dice, Hausdorff distance (HD) and Surface distance scores per model during cross-validation for the gland segmentation task. Each segment contains a pair of boxplots, where the left one corresponds to the results of the model on the full data, and the right one on the cropped data.

Figure 5

Segmentations of the prostate gland. Column (A) contains the volumes, column (B) contains the ground truth, column (C) contains the nnU-Net segmentations and column (D) contains the d2aunet segmentations. Rows are interleaved, showing a full volume and a model cropped by the object detection model, respectively.

Figure 6

Boxplots showing the distribution of Dice, Hausdorff distance (HD) and Surface distance scores per model during cross-validation for the transition segmentation task. Each segment contains a pair of boxplots, where the left one corresponds to the results of the model on the full data, and the right one on the cropped data.

Figure 7

Segmentations of the transition zone. Column (A) contains the volumes, column (B) contains the ground truth, column (C) contains the nnunet segmentations and column (D) contains the aunet segmentations. Rows are interleaved, showing a full volume and a model cropped by the object detection model, respectively.

Figure 8

Boxplots showing the distribution of Dice, Hausdorff distance (HD) and Surface distance scores per model during cross-validation for the peripheral segmentation task. Each segment contains a pair of boxplots, where the left one corresponds to the results of the model on the full data, and the right one on the cropped data.

Figure 9

Segmentations of the peripheral zone. Column (A) contains the volumes, column (B) contains the ground truth, column (C) contains the nnU-Net segmentations and column (D) contains the highResNet segmentations. Rows are interleaved, showing a full volume and a model cropped by the object detection model, respectively.