[Paper Review] A Novel Mask R-CNN Model to Segment Heterogeneous Brain Tumors through Image Subtraction

Sanskriti Singh|arXiv (Cornell University)|Apr 4, 2022

Brain Tumor Detection and Classification6 citations

TL;DR

This paper proposes a novel Mask R-CNN model enhanced with image subtraction to improve segmentation of heterogeneous brain tumors in MRI scans. By subtracting pre-contrast (T1) from post-contrast (T1gd) images before feeding them into a ResNet-18 backbone pre-trained on RSNA pneumonia data, the model achieves a DICE coefficient of 0.75 on the BraTS2020 test set, outperforming the baseline without subtraction (0.69).

ABSTRACT

The segmentation of diseases is a popular topic explored by researchers in the field of machine learning. Brain tumors are extremely dangerous and require the utmost precision to segment for a successful surgery. Patients with tumors usually take 4 MRI scans, T1, T1gd, T2, and FLAIR, which are then sent to radiologists to segment and analyze for possible future surgery. To create a second segmentation, it would be beneficial to both radiologists and patients in being more confident in their conclusions. We propose using a method performed by radiologists called image segmentation and applying it to machine learning models to prove a better segmentation. Using Mask R-CNN, its ResNet backbone being pre-trained on the RSNA pneumonia detection challenge dataset, we can train a model on the Brats2020 Brain Tumor dataset. Center for Biomedical Image Computing & Analytics provides MRI data on patients with and without brain tumors and the corresponding segmentations. We can see how well the method of image subtraction works by comparing it to models without image subtraction through DICE coefficient (F1 score), recall, and precision on the untouched test set. Our model performed with a DICE coefficient of 0.75 in comparison to 0.69 without image subtraction. To further emphasize the usefulness of image subtraction, we compare our final model to current state-of-the-art models to segment tumors from MRI scans.

Motivation & Objective

• To improve the accuracy of brain tumor segmentation in heterogeneous MRI scans using deep learning.
• To address the challenge of variable tumor appearance across T1, T1gd, T2, and FLAIR sequences in clinical practice.
• To evaluate the impact of image subtraction as a preprocessing step on segmentation performance.
• To compare the proposed model against baseline and state-of-the-art methods using standard metrics like DICE, precision, and recall.

Proposed method

• The model uses Mask R-CNN with a ResNet-18 backbone pre-trained on the RSNA pneumonia detection dataset for feature extraction.
• Image subtraction is applied between T1 and T1gd MRI sequences to highlight contrast-enhancing tumor regions.
• The subtracted image pair is used as input to the Mask R-CNN model for instance segmentation of brain tumors.
• The model is trained and evaluated on the BraTS2020 dataset, which includes multi-sequence MRI scans and expert-annotated segmentations.
• The training process leverages transfer learning and fine-tuning on the BraTS2020 validation and test splits.
• Performance is evaluated using standard metrics: DICE coefficient, precision, and recall on the untouched test set.

Experimental results

Research questions

• RQ1 Does image subtraction between T1 and T1gd MRI sequences improve tumor segmentation performance in a Mask R-CNN framework?
• RQ2 How does the proposed model compare to a baseline Mask R-CNN without image subtraction in terms of DICE score and segmentation metrics?
• RQ3 Can the combination of image subtraction and transfer learning from a pre-trained model enhance generalization on heterogeneous brain tumor segmentation?
• RQ4 How does the proposed method perform relative to state-of-the-art models on the BraTS2020 benchmark?

Key findings

• The proposed model achieved a DICE coefficient of 0.75 on the BraTS2020 test set, representing a significant improvement over the baseline model without image subtraction, which scored 0.69.
• The model demonstrated higher precision and recall compared to the baseline, indicating more accurate and complete tumor boundary detection.
• Image subtraction effectively enhanced tumor contrast, enabling the model to better distinguish enhancing tumor regions from surrounding tissue.
• The use of transfer learning from the RSNA pneumonia detection dataset improved feature learning and model convergence on the brain tumor segmentation task.
• The model outperformed the baseline in all evaluation metrics, validating the effectiveness of the image subtraction preprocessing step.
• The results position the proposed method as competitive with state-of-the-art approaches in the context of BraTS2020.