💻 computer science

MIS-HCC: Hierarchical Channel Clustering for Efficient Medical Image Segmentation

The paper proposes MIS-HCC, a hierarchical channel clustering method that utilizes Wasserstein distance to fuse similar channels in pre-trained medical image segmentation models, achieving superior accuracy and compression efficiency compared to state-of-the-art techniques for deployment on resource-limited platforms.

Original authors: Bo Zhao, Haoran Yu, Lifei Liu, Zongcheng Chu, Yining Liu, Chang Liu, Szu-Yu Chen, Zequn Xie

Published 2026-07-21

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Bo Zhao, Haoran Yu, Lifei Liu, Zongcheng Chu, Yining Liu, Chang Liu, Szu-Yu Chen, Zequn Xie

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 a world where your smartphone or a tiny medical device could instantly spot a tumor in an X-ray or highlight a broken bone in a scan, all without needing a massive, power-hungry supercomputer. This is the dream of "edge AI"—putting smart, complex computer brains into small, resource-limited gadgets. But here's the catch: the smartest medical AI models today are like giant, heavy elephants. They are incredibly accurate but require so much memory and processing power that they simply can't fit on a mobile phone or a portable scanner. To fix this, scientists have been trying to "prune" these models, which means cutting out parts they think are useless to make them smaller and faster. However, the old way of pruning is a bit like a clumsy gardener who just chops off random branches without looking at how the tree grows. They might accidentally cut off a branch that was holding up a crucial fruit, leaving the model confused and less accurate. The big question is: how do we shrink these giant AI brains without breaking the parts that actually help them see?

This paper introduces a clever new gardening tool called MIS-HCC (Hierarchical Channel Clustering for Efficient Medical Image Segmentation). Instead of just randomly chopping away parts of the AI, the researchers decided to look for twins. In these complex AI networks, there are many "channels" (think of them as individual workers or sensors) that are doing almost the exact same job. The old methods would just delete the ones they thought were "lazy," but this new method realizes that if two workers are doing the same thing, you don't need to fire one; you can just merge them into one super-worker who does both jobs at once.

To find these twins, the researchers used a special mathematical ruler called the Wasserstein distance. Imagine you have two bags of marbles, and you want to know how similar they are. A simple ruler might just count the colors, but the Wasserstein distance is like a smart mover that calculates exactly how much effort it would take to rearrange the marbles in one bag to look exactly like the other. This helps the AI understand that two channels might look slightly different on the surface but are actually carrying the same deep meaning. Once the AI finds these similar groups of channels, it uses a technique called hierarchical clustering to group them together, like organizing a messy closet by matching similar shirts and then folding them into a single, neat pile.

The results of this "merge, don't delete" strategy are impressive. The team tested their method on three different medical image datasets: breast ultrasound images, cell nuclei, and skin lesions. They also tried it on a massive 3D dataset of abdominal organs. When they compressed the models by removing 87.5% of the original channels (a huge reduction!), the old methods often lost their accuracy, making mistakes like blurring the edges of a tumor or missing a small organ entirely. In contrast, the new MIS-HCC method kept the AI sharp. For example, on the breast ultrasound dataset, while other methods saw their accuracy drop significantly, MIS-HCC performed slightly better than the other compressed methods and remained highly comparable to the original, uncompressed model, achieving a Dice score of 0.7841 compared to the original 0.7806. Even on the tricky 3D abdominal scans, where other methods struggled to keep the organs distinct, MIS-HCC stayed incredibly close to the original performance, losing only 0.14% in accuracy.

The paper suggests that by respecting the natural connections between different parts of the AI and fusing similar features rather than just deleting them, we can create medical AI that is both tiny enough to run on a phone and smart enough to save lives. It's not just about making things smaller; it's about making them smarter by keeping the right parts together. The authors believe this approach could be a key step toward bringing powerful medical diagnostics to places where big computers can't go, though they note that more testing in real-world scenarios is needed to see how it holds up in the wild.

Technical Summary: MIS-HCC for Efficient Medical Image Segmentation

Problem Statement

The deployment of complex deep neural networks (e.g., U-Net, Transformers) for medical image segmentation faces a critical bottleneck: the trade-off between high accuracy and the computational/parameter constraints of resource-limited medical platforms (e.g., mobile devices, edge nodes). While existing model compression techniques such as network pruning, quantization, and knowledge distillation exist, they often fail to address the specific characteristics of medical imagery.

Current structured pruning methods (e.g., HRank, CHEX) typically evaluate network units in isolation, overlooking the intrinsic inter-channel correlations and high feature similarity induced by anatomical structures. Consequently, these methods risk discarding essential fine-grained semantics by simply deleting "redundant" filters, rather than preserving the collective representational power of correlated features. There is a need for a compression framework that integrates representation similarity to fuse correlated features rather than merely eliminating them.

Methodology: MIS-HCC

The authors propose MIS-HCC (Hierarchical Channel Clustering for Efficient Medical Image Segmentation), a structured compression framework designed to replace binary "keep-or-drop" decisions with intelligent parameter fusion. The method operates in three primary stages:

Representation Similarity via Wasserstein Distance:
Instead of using standard metrics like cosine similarity on raw vectors, MIS-HCC employs the Wasserstein distance to quantify the similarity between channel distributions within a pre-trained network layer. This metric is chosen for its ability to capture distributional discrepancies and its robustness to local mass shifts, making it suitable for channel features aggregated over spatial, temporal, or kernel dimensions.
- For a layer with input channels, the weight tensor is flattened to form a channel feature matrix.
- A pairwise channel distance matrix is computed using the Wasserstein distance.
- This distance matrix is normalized and converted into a similarity matrix ( $M_2$ ) where higher values indicate greater similarity.
Hierarchical Clustering:
To account for the interdependence of channels, the method applies agglomerative hierarchical clustering with Average Linkage on the derived similarity matrix.
- Channels are initially treated as individual clusters.
- The algorithm iteratively merges clusters based on the average distance between their members.
- This process partitions the $N_{in}$ input channels into $K$ clusters ( $K \leq N_{in}$ ), determined by a desired compression ratio or dendrogram cut threshold. This approach ensures that functionally overlapping channels are grouped together.
Parameter Fusion:
Once channels are clustered, the method fuses the parameters within each cluster to create a compact network.
- For each cluster $c$ with index set $I_c$ , the weights are fused by arithmetic averaging along the input-channel dimension.
- This results in a new weight tensor with $K$ input channels, effectively reducing the model size and FLOPs by a factor of $K/N_{in}$ .
- Crucially, this fusion is a closed-form, training-free operation that maintains the original network topology, ensuring compatibility with general-purpose hardware and avoiding the need for specialized inference kernels or retraining.

The framework is designed to be applicable to both CNN and Transformer backbones for 2D and 3D data.

Key Contributions

The paper outlines three primary contributions:

MIS-HCC Framework: A structured compression method that introduces hierarchical clustering to medical model pruning, shifting from simple filter removal to intelligent parameter fusion.
Wasserstein-Based Similarity: The utilization of Wasserstein distance to construct a robust channel similarity matrix, explicitly modeling inter-channel correlations to preserve critical visual features during compression.
Empirical Validation: Extensive experiments on multiple medical segmentation benchmarks demonstrating that MIS-HCC achieves superior compression-accuracy trade-offs compared to state-of-the-art pruning methods.

Experimental Results

The method was evaluated on three 2D datasets (BUSI, DSB, ISIC) using U-Net and U-Net++ backbones, and one 3D dataset (FLARE 2021) using the MedSAM2 backbone. A uniform pruning rate of 87.5% was applied for comparison.

2D Segmentation: On the BUSI, DSB, and ISIC datasets, MIS-HCC outperformed existing methods (RP, PFEC, HRank, CHEX, TPP) in both Dice and IoU scores. Notably, while other methods often showed accuracy drops relative to the baseline, MIS-HCC frequently achieved scores slightly higher than the uncompressed baseline or maintained near-lossless performance. For instance, on the BUSI dataset with U-Net, MIS-HCC achieved a Dice score of 0.7841 compared to 0.7781 for the next best method (TPP).
3D Segmentation: On the FLARE 2021 dataset with MedSAM2, MIS-HCC significantly outperformed other strategies. While random pruning (RP) caused a 10.35% drop in Dice and TPP caused a 1.61% drop, MIS-HCC reduced the Dice score by only 0.14% (0.8471 vs. 0.8483 baseline).
Compression Efficiency: At high pruning rates (87.5%), MIS-HCC demonstrated superior reduction in FLOPs and parameters compared to TPP while maintaining higher accuracy.
Ablation Studies: Experiments confirmed that combining Wasserstein distance with Hierarchical Clustering yields the best performance, outperforming variants that use only one of these components or standard metrics like Cosine similarity.
Visual Analysis: Qualitative results showed that MIS-HCC produced segmentation maps with sharper lesion contours and better preservation of fine details compared to other methods, which often exhibited blurred edges or false positives.

Significance and Claims

The paper claims that MIS-HCC offers an effective solution for deploying medical image segmentation models on resource-limited platforms. By leveraging the intrinsic structural similarities prevalent in medical imagery, the method successfully balances model compression with the preservation of critical representational power.

The authors assert that their approach is particularly effective for large-scale 3D segmentation tasks where feature diversity is crucial. They position MIS-HCC as a robust and generalizable framework applicable to various architectures (CNNs and Transformers) and data modalities (2D and 3D), providing a promising path forward for practical clinical deployment without the heavy computational burdens of current state-of-the-art models. The work concludes by suggesting that these principles could be extended to other real-world scenarios to further improve model efficiency.