This repository hosts the source code and resources for the project Exploring CNN Components for Tumor Segmentation in MRI Images: An Ablation Study.
Figure 1: Proposed segmentation network,Created by author.
The dataset, sourced from kaggle, consists of 2146 images with tumors annotated in COCO Segmentation format. It's part of the TumorSeg Computer Vision Project, which focuses on Semantic Segmentation and aims to accurately identify tumor regions within Medical Images using advanced techniques
The project focuses on evaluating different configurations of an encoder-decoder convolutional neural network (CNN) to identify the most effective techniques for tumor segmentation using 2D MRI brain tumor images.
Figure 2: CNN architecture for brain tumor segmentation task. Created by author.
To study the effects of network structure on segmentation results, I conducted ablation studies on different parts of my base network (see Figure 1). All experiments were performed on Google Colab with GPU acceleration, and the learning rate was adjusted based on the learning curve of each experiment.
I used standard metrics such as IoU, Dice, precision, and recall to evaluate the performance of each model. For an explanation of these metrics, please refer to this Medium article. I use a threshold of 0.5 to generate the prediction masks from the probability maps.
In the final experiment, I incorporated the findings from the previous ablation studies. I used the baseline model with dilated convolutions as explained earlier, applied the BCE loss function, and randomly applied augmentation to 50% of the samples. Table 5 shows the segmentation results on the validation and test sets. Compared to Table 4, the test results show a significant improvement over the previous experiments.
Figure 3: CNN architecture for brain tumor segmentation task. Created by author.
Figure 4: CNN architecture for brain tumor segmentation task. Created by author.
ResNet-18 showed better performance compared to VGG-16.
Binary Cross Entropy achieved slightly better training results than the Dice coefficient.
Incorporation of dilated convolutions in the decoder significantly enhanced segmentation accuracy.
Data augmentation was used to increase the model's robustness and generalizability.
In this study, I proposed an encoder-decoder CNN for tumor semantic segmentation and studied the impacts of different network components on its performance. Based on my findings, I suggest a configuration that achieves an IoU of 0.51 and Dice of 0.63 on the test set.