💻 computer science

ConvNeXt-FD: A Fractal-Based Deep Model for Robust Biomedical Image Segmentation

This paper introduces ConvNeXt-FD, a novel deep learning architecture that combines a ConvNeXt backbone with a fractal dimension-inspired boundary-aware loss function to achieve robust and accurate segmentation across six diverse biomedical imaging datasets.

Original authors: Joao Batista Florindo, Amanda Pontes de Oliveira Ornelas

Published 2026-05-22

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Joao Batista Florindo, Amanda Pontes de Oliveira Ornelas

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 trying to trace the outline of a cloud in the sky, but the cloud is made of fog, the wind is blowing, and sometimes the edges are just a little blurry. Now, imagine doing this for thousands of different clouds, some of which are tiny specks and others are huge, messy blobs. This is essentially what doctors face when they look at medical images like ultrasounds or X-rays to find tumors, cells, or organs. The "fog" is the noise and low contrast in the images, and the "blurry edges" make it hard to know exactly where the disease starts and stops.

This paper introduces a new digital tool called ConvNeXt-FD to help solve this tracing problem. Here is how it works, broken down into simple concepts:

1. The Smart Tracer (The Architecture)

Think of the software as a very skilled artist who uses a specific type of brush.

The Backbone (ConvNeXt): The authors chose a modern, powerful "brush" called ConvNeXt. It's like a super-charged version of the traditional tools used in the past. It's designed to look at an image and understand both the big picture (like "this is a lung") and the tiny details (like "this is a tiny crack in the tissue").
The Blueprint (U-Net): They built this brush into a familiar structure called a U-Net. Imagine a funnel that squeezes the image down to understand its core meaning, and then a funnel that expands it back out to draw the final picture. As it expands, it grabs notes it took earlier (skip connections) to make sure the final drawing is sharp and detailed.

2. The "Fractal" Compass (The Loss Function)

This is the paper's most creative idea. Usually, when a computer tries to draw a shape, it just checks: "Did I get the pixels right?" But that's not enough if the edge is jagged or fuzzy.

The authors added a special rule to the computer's training, inspired by Fractals.

The Analogy: Imagine you are trying to trace the edge of a coastline. A simple ruler might say, "It's 10 miles long." But if you look closer, the coastline is full of tiny bays and rocks. A fractal is a way of measuring how "rough" or "complex" that edge is.
The Application: The computer doesn't just learn to draw the shape; it also learns to measure the "roughness" of the edge. It has a side-brain that constantly asks, "Does the jaggedness of my drawing match the jaggedness of the real object?" If the computer draws a smooth line where the real object is rough, it gets penalized. This forces the AI to pay extra attention to the messy, fuzzy boundaries where diseases often hide.

3. The Training Ground (The Datasets)

To prove this new tool works, the authors tested it on six different "training grounds," each with its own unique difficulty:

BUSI & DDTI: Looking for lumps in breast and thyroid ultrasounds (very foggy images).
FluoCells: Counting glowing cells under a microscope (often clumped together).
IDRiD: Finding the optic disc in eye scans (very precise circles).
ISIC2018: Tracing skin lesions (irregular shapes).
MoNuSeg: Separating individual cell nuclei that are packed tightly like grapes.

4. The Results (What Happened)

The paper claims that ConvNeXt-FD performed exceptionally well, often beating the current best tools.

The Secret Sauce: The biggest boost in performance came from pre-training. Imagine teaching the artist to draw by first showing them millions of photos of cats, cars, and trees (ImageNet) before showing them medical images. The authors found that starting with this general knowledge made the AI much better at understanding medical images, especially for the foggy ultrasound scans where it improved accuracy by over 16%.
The Fractal Effect: The "roughness" rule (the Fractal loss) was crucial. For skin lesions, which have very wild, irregular edges, the computer needed a strong version of this rule to get it right. For smoother shapes, a lighter touch worked better.

Summary

In short, the authors built a new AI model that combines a powerful modern "brain" (ConvNeXt) with a special "compass" that measures the complexity of edges (Fractal Dimension). By teaching this model to care about how rough and detailed the boundaries are, and by giving it a head start with general image knowledge, they created a tool that can trace medical shapes more accurately than many existing methods, even in the messiest, most confusing images.

Technical Summary: ConvNeXt-FD for Robust Biomedical Image Segmentation

Problem Statement
Biomedical image segmentation is essential for diagnosis and treatment planning but faces significant challenges due to inherent image variability, noise, intensity inhomogeneities, and complex morphologies. While deep learning architectures, particularly U-Net variants and emerging Vision Transformers, have advanced the field, achieving robust precision at object boundaries remains difficult. Traditional loss functions, such as Binary Cross-Entropy (BCE) and standard Dice loss, often struggle with small objects, class imbalance, and fuzzy boundaries, leading to suboptimal shape fidelity in clinical contexts.

Methodology
The paper proposes ConvNeXt-FD, a novel deep learning framework designed to address these limitations through a hybrid architecture and a specialized loss function.

Architecture: The model adopts a U-Net-like encoder-decoder structure.
- Encoder: Instead of standard CNNs or Transformers, the encoder utilizes the ConvNeXt backbone (specifically 'tiny', 'base', or 'large' variants). This design incorporates modern CNN principles, including large kernel depthwise convolutions (7×7), inverted bottleneck structures similar to Transformer MLP blocks, and Layer Normalization. The encoder extracts hierarchical features, producing feature maps at progressively smaller spatial resolutions.
- Decoder: The decoder reconstructs the segmentation mask by progressively upsampling features and integrating high-level semantic information with fine-grained spatial details via skip connections from the encoder.
- Auxiliary Branch: To support the boundary-aware loss, an auxiliary branch predicts a Fractal Dimension (FD) map.
Hybrid Loss Function: The core innovation is a hybrid loss function ( $L$ ) that combines a standard segmentation loss with a boundary-aware regularization term inspired by Fractal Geometry:
$L = L_{Dice} + \lambda_{FD} L_{Boundary}$
- $L_{Dice}$ : The standard Dice loss ensures pixel-wise overlap accuracy, particularly effective for imbalanced datasets.
- $L_{Boundary}$ : This term is derived from a differentiable formulation of Fractal Dimension. Using a differential box-counting method, the model estimates the fractal dimension of the input image to create a ground truth FD map. The loss is defined as the Mean Squared Error (MSE) between the ground truth FD map and the FD map predicted by the auxiliary branch. This approach aims to minimize the discrepancy between predicted and ground truth object shapes, explicitly encouraging the model to learn texture complexity and preserve boundary irregularity.

Key Contributions

Novel Architecture: The integration of the ConvNeXt backbone into a U-Net framework for biomedical segmentation, leveraging the inductive biases of modern CNNs for efficient feature extraction.
Fractal-Inspired Regularization: The introduction of a boundary-aware loss term based on Fractal Dimension. Unlike standard geometric losses, this method quantifies texture complexity and irregularity to improve shape fidelity, particularly in regions with fuzzy or complex boundaries.
Comprehensive Evaluation: A rigorous evaluation across six diverse public biomedical datasets: BUSI (Breast Ultrasound), DDTI (Thyroid Ultrasound), FluoCells (Fluorescent Cells), IDRiD (Optic Disc), ISIC2018 (Skin Lesions), and MoNuSeg (Nuclei).

Experimental Results
The authors evaluated ConvNeXt-FD using metrics including Dice Coefficient, Jaccard Index, Accuracy, Sensitivity, Specificity, and False Positive Rate. Key findings include:

Performance: The model achieved competitive and often state-of-the-art results across all six datasets. Notable performances include a Dice score of 96.19% on IDRiD (optic disc), 89.12% on ISIC2018 (skin lesions), and 85.10% on FluoCells.
Transfer Learning: Initializing the ConvNeXt encoder with ImageNet pre-trained weights proved critical. Datasets with subtle boundary challenges (e.g., BUSI and DDTI) showed performance improvements of over 16% in Dice scores when using pre-trained weights compared to training from scratch.
Hyperparameter Sensitivity: The weighting factor $\lambda_{FD}$ for the fractal loss was found to be task-dependent. For instance, skin lesion segmentation (ISIC2018) benefited from a higher weight ( $\lambda_{FD}=0.50$ ), while ultrasound datasets performed better with a milder regularization ( $\lambda_{FD}=0.01$ ).
Comparison: ConvNeXt-FD outperformed or matched various state-of-the-art methods, including U-Net++, TransUNet, Swin-Unet, and specialized medical segmentation networks, demonstrating robustness against speckle noise and ill-defined boundaries.

Significance and Claims
The paper claims that ConvNeXt-FD provides a strong, versatile solution for biomedical image segmentation by effectively balancing global contextual understanding with local boundary precision. The authors assert that the integration of the ConvNeXt encoder with fractal-inspired boundary regularization successfully captures both high-level semantic features and fine-grained boundary details. This leads to more accurate and clinically relevant segmentations, particularly in challenging contexts where shape fidelity is paramount. The work highlights the potential of combining modern CNN architectures with geometrically inspired loss functions to overcome the limitations of current segmentation paradigms.

Limitations and Future Directions
The authors acknowledge that the current boundary-aware loss involves a numerical estimation of the Fractal Dimension. Future work is suggested to explore more direct, differentiable formulations of shape-based metrics (e.g., from optimal transport or geometric deep learning) and adaptive strategies for $\lambda_{FD}$ . Additionally, the authors propose extending the framework to multi-modal data integration, instance segmentation, computational efficiency improvements, and uncertainty quantification.