💻 computer science

Benchmarking MRI Representations for Deep Learning-Based Focal Cortical Dysplasia Segmentation

This study systematically benchmarks various MRI representations for deep learning-based Focal Cortical Dysplasia segmentation using nnU-Net, revealing that while FLAIR is the strongest single modality, a four-channel multimodal configuration combining conventional T1w/FLAIR with ratio-derived representations achieves the highest performance with a 5.0% relative improvement in Dice score.

Original authors: Soumen Ghosh, John Phamnguyen, Amit Soni Arya, Subhojit Mandal, Tilottama Goswami, Rajat Vashistha

Published 2026-07-20

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Soumen Ghosh, John Phamnguyen, Amit Soni Arya, Subhojit Mandal, Tilottama Goswami, Rajat Vashistha

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are a detective trying to solve a mystery inside a complex, foggy city. The city is the human brain, and the mystery is a tiny, tricky patch of trouble called a "focal cortical dysplasia" (FCD). This isn't a giant, obvious monster; it's a subtle glitch in the brain's wiring that can cause seizures, making life very hard for people with epilepsy. To find this glitch, doctors use a special kind of camera called an MRI. Think of an MRI like a high-tech scanner that takes pictures of the brain's structure. But here's the catch: the glitch is so sneaky that sometimes the camera just can't see it clearly, even with its best lenses.

For years, scientists have been trying to build super-smart computer programs (called "deep learning") to act as extra eyes for these detectives. These programs are like students who study thousands of brain scans to learn how to spot the trouble spots automatically. Usually, when these students get better at their job, we assume it's because we gave them a smarter brain or a better way to study. But what if the problem wasn't the student's brain, but the textbook they were reading? What if the way we show the pictures to the computer—the "representation"—is the real key to solving the mystery? This paper asks a simple but powerful question: Does changing how we present the brain images to the computer help it find the hidden glitches better, even if the computer's brain stays exactly the same?

The researchers in this study decided to put this idea to the test using a very popular and reliable computer program called "nnU-Net." They didn't try to build a new, fancier computer brain; instead, they kept the brain exactly the same and changed the "textbooks" (the MRI images) they fed into it. They used a dataset of brain scans from 85 people with the condition and 25 healthy people to act as a control group. They tested eight different ways to show the images: using just one type of scan, mixing two types, and even creating "ratio" images. A ratio image is like taking two photos of the same scene and dividing the numbers of one by the other to highlight specific differences, kind of like using a filter to make a specific color pop out.

Here is what they found. First, they discovered that not all pictures are created equal. When they used just one type of scan, the "FLAIR" image was the clear winner, finding the trouble spots much better than the standard "T1w" image. It's as if the FLAIR camera was wearing special glasses that made the invisible visible, while the T1w camera was just looking at the normal architecture of the city. Interestingly, they found that using the "ratio" images (the divided photos) alone was a bad idea. If you tried to solve the mystery using only the ratio pictures, the computer got confused and missed most of the clues. It was like trying to read a map that only showed the shadows of the buildings but not the buildings themselves.

However, the real magic happened when they mixed things up. When they gave the computer the standard FLAIR and T1w images plus the ratio images, the performance got even better. The best combination was a "four-channel" setup, where the computer looked at the T1w image, the FLAIR image, and both ratio versions all at once. This super-team approach improved the computer's ability to draw the outline of the trouble spot by about 5% compared to just using the standard two images. While this might sound like a small number, in the world of spotting tiny, hidden brain glitches, it's a significant step forward.

The study also showed that these ratio images didn't necessarily help the computer find more trouble spots (it still missed about a third of them, just like the standard method), but they did help the computer draw the edges of the spots much more accurately once it found them. It's like the standard method could say, "Hey, there's a problem over there," but the ratio-enhanced method could say, "There's a problem over there, and here is exactly where it starts and stops."

One of the most important takeaways is that the researchers proved you don't always need to invent a brand-new, complicated computer brain to get better results. Sometimes, just changing the way you show the data—optimizing the "representation"—can do the heavy lifting. They found that the order of the ratio mattered, too; dividing the FLAIR image by the T1w image worked better than the other way around, suggesting that the way you mix the ingredients changes the flavor of the information the computer receives.

In the end, this paper suggests that while we often obsess over building bigger and smarter AI models, we might be ignoring a simpler, cheaper, and very effective tool: better image preparation. By treating the way we present MRI data as a crucial part of the solution, rather than just a boring first step, we can help our digital detectives solve the mystery of epilepsy a little more effectively. The researchers admit that this was tested on a specific set of data and that more testing is needed to see if it works for everyone, but the results are promising enough to suggest that future AI systems should pay just as much attention to their "textbooks" as they do to their "brains."

Technical Summary: Benchmarking MRI Representations for Deep Learning-Based Focal Cortical Dysplasia Segmentation

Problem Statement

Focal cortical dysplasia (FCD) is a leading structural cause of drug-resistant focal epilepsy. Accurate identification and delineation of FCD lesions via magnetic resonance imaging (MRI) are critical for presurgical planning. However, FCD lesions are often subtle and heterogeneous, making them difficult to detect on conventional MRI; approximately 30% are overlooked during standard interpretation. While deep learning, particularly U-Net architectures like nnU-Net, has improved automated segmentation, most research focuses on optimizing network architectures and training strategies. Consequently, the specific contribution of MRI representation design—how input data is preprocessed and combined—remains poorly understood. It is unclear whether ratio-derived representations (e.g., T1w/FLAIR) offer complementary value over conventional contrasts or if they can serve as standalone inputs.

Methodology

The study presents a systematic benchmark to isolate the impact of MRI representation on segmentation performance, strictly controlling for all other variables.

Dataset: The authors utilized a publicly available dataset comprising 85 subjects with FCD and 25 healthy controls. All subjects had T1-weighted (T1w) and fluid-attenuated inversion recovery (FLAIR) scans. Expert-manually delineated lesion masks served as ground truth.
Preprocessing Pipeline: A standardized pipeline was applied to all data, including affine registration (ANTs), skull stripping, N4 bias field correction, resampling to $1 \times 1 \times 1 \text{mm}^3$ isotropic resolution, and intensity normalization.
Input Representations: Eight distinct input configurations were evaluated:
1. Single Modality: T1w, FLAIR.
2. Ratio-Derived: T1w/FLAIR and FLAIR/T1w (computed voxel-wise).
3. Multimodal Fusion:
  - Conventional: T1w + FLAIR.
  - Augmented: Conventional + T1w/FLAIR; Conventional + FLAIR/T1w; Conventional + Both Ratios.
Framework: All experiments employed the nnU-Net v2 framework with a 3D full-resolution configuration. To ensure a fair comparison, the network architecture, preprocessing, data augmentation, optimization strategy (Dice + Cross-Entropy loss), and five-fold cross-validation splits were kept identical across all eight configurations.
Evaluation: Performance was measured using the Dice Similarity Coefficient (Dice), sensitivity, and precision. Metrics were reported for the entire cohort (including missed lesions) and for successfully detected lesions only. Statistical significance was assessed using the Friedman test and Wilcoxon signed-rank tests with False Discovery Rate (FDR) correction.

Key Contributions

First Systematic Benchmark: The study provides the first controlled benchmark of conventional versus ratio-derived MRI representations for deep learning-based FCD segmentation.
Isolation of Representation Effects: By fixing the nnU-Net architecture and training protocol, the authors demonstrate that performance gains can be attributed solely to input representation design rather than architectural complexity.
Analysis of Asymmetric Ratios: The work investigates the specific utility of asymmetric ratio images (T1w/FLAIR vs. FLAIR/T1w), revealing that the ordering of the ratio operation significantly impacts the information presented to the network.

Results

Single-Modality Performance: Among single inputs, FLAIR achieved the strongest performance (Dice: $0.369 \pm 0.303$ ), significantly outperforming T1w ( $0.246 \pm 0.290$ ) and the T1w/FLAIR ratio ( $0.128 \pm 0.245$ ). FLAIR also yielded the lowest lesion miss rate (28%).
Multimodal Performance: The four-channel configuration (E8: T1w + FLAIR + T1w/FLAIR + FLAIR/T1w) achieved the highest overall Dice score of 0.376, representing a 5.0% relative improvement over the conventional T1w+FLAIR baseline (E5, Dice 0.358).
Lesion Delineation vs. Detection: When analyzing only successfully detected lesions, the performance gap widened. The E8 configuration achieved a Dice score of 0.571, compared to 0.534 for the baseline. This indicates that ratio-derived representations primarily enhance lesion delineation (boundary accuracy) rather than the initial detection of the lesion.
Ratio Directionality: The FLAIR/T1w ratio consistently outperformed the T1w/FLAIR ratio. The authors attribute this to the fact that FCD lesions are characterized by FLAIR hyperintensity; placing FLAIR in the numerator preserves these signal abnormalities while normalizing anatomical variations.
Statistical Significance: While multimodal configurations with ratios consistently showed higher mean Dice scores than the baseline, these improvements were not statistically significant after multiple-comparison correction (adjusted $p > 0.05$ ). However, the Friedman test confirmed a significant overall effect of MRI representation on performance ( $p < 0.001$ ).
Case-Specific Gains: Qualitative analysis revealed substantial benefits for specific challenging cases. For instance, in one subject (FCD_072), the Dice score improved from 0.002 (baseline) to 0.814 when using ratio-enhanced inputs.

Significance and Claims

The paper argues that MRI representation design is an important, yet underexplored component of deep learning-based medical image segmentation. The authors claim that:

Optimizing input representations is a viable strategy for improving segmentation performance without increasing model complexity or altering the underlying network architecture.
Ratio-derived representations should be viewed as complementary to conventional MRI sequences rather than replacements. They provide additional contrast information that refines lesion boundaries, particularly for subtle FCD lesions.
Future research should jointly optimize MRI representation design and network architecture, rather than focusing exclusively on architectural innovation.

The study concludes that while the average improvements were modest and not statistically significant across the entire cohort, the representation design offers a computationally efficient pathway to enhance segmentation quality for difficult clinical cases.