Core Concepts
This study proposes a post-processing refinement mechanism to enhance the interpretability and robustness of deep learning model explanations for brain tumor detection from MRI images.
Abstract
This study addresses the challenge of lack of explainability in deep learning models used for medical image analysis, particularly in the context of brain tumor detection from MRI scans. The authors employ the LIME (Local Interpretable Model-agnostic Explanations) library and LIME image explainer to generate explanations for the model's predictions.
To improve the interpretability of these explanations, the authors introduce a post-processing refinement mechanism that leverages image morphology operations and heuristic rules. The key steps are:
- Detecting the brain region in the input image using edge detection techniques like Canny, Laplace, and Otsu's thresholding. This generates a binary brain mask.
- Retaining only the segments from the LIME explanation that have a significant overlap (80% or more) with the brain mask, and setting the importance of other segments to 0.
- Evaluating the refined explanations using metrics like tumor segment coverage and brain segment coverage to find the optimal number of segments (3) that balances interpretability and specificity.
The proposed refinement mechanism demonstrates significant improvements in the interpretability and accuracy of the explanations compared to the original LIME outputs. However, the authors also acknowledge the potential inconsistencies in the brain mask generation as a limitation that requires further investigation.
Overall, this work contributes to the ongoing efforts to enhance the transparency and trustworthiness of deep learning models in medical image analysis, which is crucial for their successful integration into clinical decision-making processes.
Stats
The dataset used in this study consists of 4,602 MRI images of the brain, categorized based on the presence or absence of a brain tumor.
The authors preprocessed the dataset by resizing the images to 224x224 pixels, normalizing the pixel values, and removing duplicate images, resulting in a final dataset of 4,015 images.
A Stratified K-Fold validation strategy with 5 splits was used to ensure a robust evaluation of the deep learning models' performance.
Quotes
"One major issue with these models is their lack of explainability. Because deep neural networks are complex, their decision-making processes are frequently transparent, which makes it difficult for medical experts to understand and accept the outcomes."
"Understanding the difficulties in obtaining results that are transparent, we use an explainability method that is specific to the complexities of medical image analysis."
"To be more specific, after the use of the VGG Image Annotator, a new mask that represents the location of the tumor is created, call it Tum. Meaning that a pixel of the original image (x, y) belongs in the Tumor Mask, if-f this pixel is inside of the tumor polygon that is created."