Geometry-Aware Superpixel Graph Transformer with Metadata for Skin Lesion Classification

arXiv:2606.20390 · cs.CV · Submitted 2026-06-18 · Read on arXiv

Listen

Transcript

Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.

Tom: Today's paper: "Geometry-Aware Superpixel Graph Transformer with Metadata for Skin Lesion Classification".

Jane: Automated skin cancer classification from dermoscopic images remains challenging due to heterogeneous lesion structure, strong intra-class variability, and subtle visual differences between benign and malignant cases.

Tom: First, who's behind it and why it matters.

Title and authors: Tom: So Jane, this paper introduces something called "Geometry-Aware Superpixel Graph Transformer with Metadata for Skin Lesion Classification," and it sounds really interesting because it tackles those tricky issues we've been seeing with skin cancer classification.

Jane: It does sound complex, Tom, but the title tells us the core idea is combining geometry, superpixels, graphs, and metadata to help classify lesions better than current methods.

Lu: What I find fascinating is how they explicitly model lesions as graphs of spatially coherent superpixel regions represented by frozen CNN features; it’s like giving the visual data a proper spatial structure instead of just flattening it into tokens or patches <ref:2606.20390#pg2>.

Meng: From an engineering standpoint, I'm wondering how they manage keeping the CNN backbone frozen while still adapting something meaningful in the graph module; that’s a big constraint for practical deployment.

Lalam: The idea of introducing a dedicated metadata context node connected to all regions is very compelling because it moves patient data from being just an auxiliary vector into being integrated directly into the relational space of the image, which could seriously improve how we interpret medical images <ref:2606.20390#pg1>.

Tom: Exactly, Jane, and that integration of metadata sounds like a big step forward because it's not just tacked on at the end; they're making it part of the core reasoning process.

Jane: It really is about moving beyond simple late fusion where you just dump patient info into the final layer; this paper suggests a more structured way for that context to influence how neighboring regions interact during classification.

Lu: The methodology involves decomposing the image into SLIC superpixels to form region nodes, and then constructing edges between these nodes based on Euclidean distance between their centroids, which gives us an explicit geometric relationship <ref:2606.20390#pg2>.

Meng: So, instead of a model looking at a patch in isolation, it’s looking at how that patch relates spatially to every other significant region in the lesion structure? That makes sense for capturing fine details.

Lalam: And they enrich those edges with geometric attributes like distance and orientation during the graph construction phase, which is crucial because those attributes tell the model *how* regions are related spatially <ref:2606.20390#pg2>.

Tom: That edge-aware attention mechanism sounds like it's really smart, allowing the model to learn which geometric relationships matter most for distinguishing benign from malignant cases.

Title and authors: Jane: The paper also describes a similarity-weighted structural refinement step where they aggregate information based on feature similarity between connected nodes, which helps ensure semantic consistency among the superpixels <ref:2606.20390#pg2>.

Lu: I think what really sets this apart is that the edge-aware graph transformer design allows attributes to actively "shut down" or "boost" the influence of specific neighbors during message passing, which is a sophisticated form of local adaptation <ref:2606.20390#pg2>.

Meng: If we can selectively emphasize informative neighbors based on geometry, that could lead to much more interpretable results, even if it makes the computation heavier initially.

Lalam: That capability to selectively boost or suppress neighbor influence means the AI system won't just rely on a general feature aggregation; it will learn to focus its attention precisely where the geometric evidence for pathology is strongest <ref:2606.20390#pg1>.

Tom: So, we’re talking about a system that doesn't just classify based on what it sees globally, but understands the precise spatial arrangement of features within the lesion structure itself.

Jane: And this moves us closer to building systems that can provide more detailed justifications for their diagnoses because they are grounded in the visual geometry <ref:2606.20390#pg1>.

Lu: The results show consistent accuracy gains across several public benchmarks, achieving scores like ninety-eight point six one percent on ISIC2024 and ninety-eight point two three percent on HAM10000, which validates the approach's effectiveness even with complex datasets <ref:2606.20390#pg1>.

Meng: Those numbers are solid, but I’m curious about the practical implications for real-world clinical workflows; can this run fast enough when we need to scan a large volume of images?

Lalam: The paper confirms that incorporating geometric edge encoding leads to substantial accuracy improvements across all datasets, showing that adding structured spatial awareness yields tangible performance gains <ref:2606.20390#pg1>.

Tom: And the metadata integration is also significant, as the study confirms its importance because it substantially improves accuracy across all benchmarks when integrated as a dedicated graph node <ref:2606.20390#pg1>.

Jane: So, to summarize, this paper proposes a framework called "Geometry-Aware Superpixel Graph Transformer with Metadata for Skin Lesion Classification" that structures the visual input into a graph and weaves patient context directly into that graph structure using edge-aware attention <ref:2606.20390#pg2>.

Lu: The implication here is that we can build models that aren't just pattern recognizers but are actually modeling the underlying spatial relationships inherent in medical images, which is a big step for understanding lesion morphology <ref:2606.20390#pg1>.

Title and authors: Meng: From an engineering perspective, confining the adaptation to the graph module while keeping the CNN frozen makes it much more feasible to deploy this kind of specialized reasoning engine in production environments <ref:2606.20390#pg1>.

Lalam: For culture, this work suggests that future AI models should prioritize building relational structures over just feature vectors because understanding *how* parts connect is a richer form of knowledge than just knowing what the parts are <ref:2606.20390#pg1>.

Tom: So we’re seeing a real push toward more structured, spatially aware reasoning in medical AI, and this paper is showing us how to achieve that by building explicit geometric relationships within the classification process <ref:2606.20390#pg2>.

Jane: It seems like the main implication is moving from models that just look at things and then check if they match a category, to models that reason about the spatial arrangement of features to determine the category itself <ref:2606.20390#pg1>.

Lu: Thinking bigger, this methodology could inspire future work in other visual domains where fine-grained structural relationships are key, perhaps even in analyzing complex biological systems <ref:2606.20390#pg1>.

Meng: I see the potential for better interpretability too; if we can trace the decision back to specific edge interactions, it helps us debug why a certain classification was made <ref:2606.20390#pg1>.

Lalam: It really shows that integrating context in a structured way, rather than just stacking it on top, fundamentally alters the model's capacity for nuanced understanding <ref:2606.20390#pg1>.

Tom: We’re wrapping up our discussion on "Geometry-Aware Superpixel Graph Transformer with Metadata for Skin Lesion Classification," a paper that successfully uses graph structures and metadata fusion to achieve high accuracy in skin cancer classification <ref:2606.20390#pg2>.

Jane: It gives us a clear path forward for building more context-aware visual reasoning systems by treating image regions as nodes in a graph and integrating patient data directly into that relational space <ref:2606.20390#pg1>.

Lu: I'm excited to see how this framework expands beyond skin lesions, especially since the concept of encoding inter-regional geometry as edge attributes is something we can apply widely across different visual data types <ref:2606.20390#pg1>.

Meng: For implementation, the key will be ensuring that the superpixel decomposition and graph construction steps are efficient enough to handle high-resolution images without crippling inference speed <ref:2606.20390#pg1>.

Lalam: The potential impact is that AI systems become inherently more capable of understanding complex visual evidence by learning the spatial grammar of those images, which could improve diagnostics across many fields <ref:2606.20390#pg1>.

The paper's summary: Tom: So, to quickly recap, this paper is all about taking those complex skin lesion images and turning them into a structured graph where every small region is a node, and they’re weaving in patient data right into that graph using special connections.

Jane: That’s exactly it; instead of just looking at the image as one big blob of pixels or a few patches, this AI treats each distinct part of the lesion as an individual piece with its own identity and spatial relationships to everything else.

Lu: And what I find really compelling is how they use those geometric details—like distance and orientation between regions—as actual attributes on the edges connecting them, which gives the model a much richer understanding of how things are arranged spatially.

Meng: From an engineering standpoint, that structured approach means we’re not just feeding raw pixels into a black box; we’re giving the AI a map of where things are and how they connect, which should lead to more reliable and interpretable results when we deploy it.

Lalam: I think the biggest cultural impact here is how this moves us toward AI systems that can provide visual justifications for their decisions, helping clinicians understand *why* a certain area looks concerning based on its geometric context.

Tom: Right! And they’re showing that by integrating patient information not just as an extra string of numbers but as a dedicated node in the graph, we get much more robust reasoning across all those datasets.

Jane: It really shows that the way we structure our data input can fundamentally change what an AI model is actually capable of learning about a problem.

Lu: The ability to selectively amplify or suppress neighbor influence using that edge-aware transformer mechanism is fascinating; it’s like the AI can focus its attention precisely on the most geometrically relevant parts of a lesion when making a classification.

Meng: That selective attention sounds like something we could use in other complex visual tasks where noise is high; it gives us a way to filter out irrelevant visual clutter based on spatial logic rather than just raw pixel intensity.

Lalam: And thinking about the broader picture, this work suggests that AI systems can evolve from simply pattern matching to actually understanding the spatial grammar of medical pathology, which could help train future models in vastly different visual domains down the line.

Tom: Exactly! We're moving beyond just seeing a picture and learning how to reason about the relationships *inside* that picture, which is a huge step for diagnostic tools.

Jane: It’s wonderful to see this kind of structured reasoning being applied to something as critical as skin cancer classification, giving us hope for more nuanced AI in medicine.

Lu: So, the method seems to be combining superpixel feature extraction with graph-native reasoning and metadata fusion into a single pipeline for top performance on those challenging benchmarks.

Meng: I’m focusing on the practical side now; while the accuracy gains are impressive, the next big hurdle will be optimizing that graph construction and message passing so this architecture can run efficiently on standard medical imaging hardware.

Lalam: And looking ahead, I think this kind of relational modeling is essential for creating truly intelligent agents that can reason about complex spatial hierarchies in any visual data they encounter.

Tom: That’s a fantastic direction to take our discussion; the potential here is so much bigger than just getting a higher accuracy number on a benchmark.

The paper's improvements: Tom: So, we’re talking about how this framework actually improves things compared to older methods; it’s not just about having a complex model, but about making the reasoning itself smarter by incorporating spatial relationships and patient context directly into the process.

Jane: Basically, the authors suggest that by explicitly building these graphs and using those edge-aware attention mechanisms, the AI can learn to prioritize which visual features and clinical facts are most important for making a correct diagnosis.

Lu: The main improvement is moving away from treating image data as a flat grid or patch set; instead, they show that modeling the spatial connectivity of superpixels allows the AI to capture fine-grained morphological details that define benign versus malignant lesions much better.

Meng: I see the implication for practical use being that this structure gives us a way to extract not just a label, but a kind of spatial proof for why it made that choice, which is something we need in clinical settings where trust in the diagnosis is paramount.

Lalam: From an AI perspective, the addition of the dedicated metadata node fundamentally alters how information flows through the network; it means patient history isn't just an afterthought but a core component of every region's reasoning process.

Tom: That’s right, and they also introduced that similarity-weighted propagation step which refines those initial feature embeddings to ensure that connected regions actually share semantic consistency before the final classification happens.

Jane: It’s like having a system where every piece of evidence has to check in with its neighbors to make sure it makes sense in the overall picture, which really sharpens the diagnostic power of the AI.

Lu: By confining all task-specific learning only to this graph module while keeping the backbone frozen, they demonstrate that you can achieve significant performance gains without having to retrain massive feature extractors from scratch every time you change your dataset.

Meng: That parameter efficiency is key for us; if we can get near SOTA results with a frozen base, it drastically lowers the computational cost and deployment friction for getting this kind of specialized reasoning into a real-world application.

Lalam: This approach opens up possibilities where AI systems can become much more capable of handling subtle visual cues that human experts might miss when they are looking at thousands of images without the benefit of explicit spatial mapping.

Tom: It’s really about making the AI's decision-making process transparent and spatially grounded, which is a huge step toward building more dependable medical diagnostic tools.

Jane: So, we’re seeing an evolution in how we teach AI to "see" and reason about complex visual data by giving it explicit rules for spatial organization and context integration.

Lu: The way they handle the edge attributes—allowing them to boost or suppress specific neighbors—is what really makes this model dynamic; it learns exactly which geometric relationships are most predictive for a certain outcome.

Meng: That ability to adapt locally based on geometric relevance is powerful, and I’m curious if we can extend that idea to other complex data modalities where spatial context matters as much as the raw features.

Lalam: The long-term vision here is an AI culture where visual understanding isn't just about pattern recognition but about building a structured, context-aware representation of reality itself.

Tom: We’re definitely seeing that shift; it’s moving from simple data processing to true relational reasoning in the eyes of the machine.

Conclusion: Tom: So, to wrap up our discussion on "Geometry-Aware Superpixel Graph Transformer with Metadata for Skin Lesion Classification," this paper shows how explicitly modeling spatial relationships and integrating clinical data directly into a graph structure leads to solid performance gains on challenging skin cancer tasks.

Jane: It really demonstrates that moving the AI from just looking at visual pixels to understanding the geometry and context between those regions is a very effective way to build more reliable systems.

Lu: The core takeaway is that structured, spatially aware reasoning can unlock capabilities in vision models that rely on complex spatial arrangements within an image, which is something I think has huge potential for other fields.

Meng: From an engineering viewpoint, the fact that they could achieve these gains while keeping the main CNN backbone frozen means we have a very viable path toward deploying this kind of specialized reasoning into existing systems without needing massive retraining cycles.

Lalam: For culture, this work shows that AI can be built to be more inherently contextual and explainable, which is vital for building trust in any system that makes critical decisions about human health.

Tom: Exactly! It’s a big step toward making AI not just accurate at classification, but also capable of providing a spatially grounded justification for its findings.

Jane: We’re really seeing how relational modeling can become the next major direction for improving visual understanding in complex domains like medicine.

Lu: I think the ability to selectively emphasize or suppress neighbor influence via that edge-aware transformer is something that could be applied across a wide spectrum of visual tasks where context filtering is essential.

Meng: That local adaptation capability sounds like it addresses a real problem with large models—how they can focus their computational effort efficiently instead of processing everything equally.

Lalam: This approach suggests that the future of impactful AI lies in building systems that can learn the underlying spatial grammar of visual data, which could improve diagnostic consistency across many different applications.

Tom: It’s been fantastic digging into the details of "Geometry-Aware Superpixel Graph Transformer with Metadata for Skin Lesion Classification" and seeing how these architectural choices translate directly into measurable accuracy improvements.

Jane: We’ve seen that combining geometry, graphs, and metadata in this structured way provides a much more robust foundation for AI reasoning than just feeding raw features into a standard classifier.

Lu: This paper really solidifies the idea that building explicit geometric relationships is not just an academic exercise but a functional way to enhance the AI's capacity for nuanced visual understanding.

Meng: I’m looking forward to seeing how the developers handle scaling this up, as optimizing that graph construction step for high-resolution medical scans will be the next major engineering challenge.

Lalam: Ultimately, this research pushes us toward a cultural shift where we expect AI not just to answer questions, but to provide spatially informed and contextually aware explanations for its answers.

Tom: And that’s exactly what we wanted to highlight about "Geometry-Aware Superpixel Graph Transformer with Metadata for Skin Lesion Classification"—it’s all about making the AI smarter by giving it a better map of the visual world.

Muhammad Azeem, Tanveer Hussain, Amr Ahmed, Ardhendu Behera

Edge Hill University

cs.CV

Submitted: 2026-06-18

Updated: 2026-06-25

Code: https://github.com/azeemchaudharyg/GeoMeta-GT

Importance score: 92/100

The gist: Automated skin cancer classification from dermoscopic images remains challenging due to heterogeneous lesion structure, strong intra-class variability, and subtle visual differences between benign

Key concepts

Superpixel Decomposition: The input image is broken down into small, coherent regions called superpixels using SLIC. Features for each region node are created by combining statistics from the entire region (global context) with statistics from the most prominent local area (local saliency). This creates rich descriptors for every part of the lesion.
Metadata-as-Node Fusion: Patient clinical data is treated as a dedicated node in the graph, separate from image regions. This metadata node connects to all superpixel nodes. By embedding this information into the region features and using a dedicated connection, the model can use demographic or clinical variables to inform how different parts of the lesion relate to each other.
Edge-aware Graph Transformer: This mechanism enhances reasoning by explicitly including edge attributes in the attention calculation. When a node looks at its neighbors, it considers not just their features but also the specific relationship (edge) between them. This allows the model to selectively emphasize or suppress connections based on task relevance.
Geometry-Attributed Graph: The lesion is represented as a graph where regions are nodes and spatial proximity defines the edges. These edges carry geometric information, such as Euclidean distance between region centroids. This structure allows the network to reason about the physical layout and spatial relationships of different lesion components.

Terminology

Transformer — 36× in this paper · explained in 2 episodes
Accuracy — 25× in this paper · explained in 1 episode
Benchmark — 14× in this paper · explained in 2 episodes
Core — 6× in this paper · explained in 2 episodes
Graph Construction — 6× in this paper · explained in 3 episodes
Ablation — 5× in this paper · explained in 1 episode
Adaptation — 5× in this paper · explained in 1 episode
Consistency — 5× in this paper · explained in 3 episodes
Aggregation — 4× in this paper · explained in 1 episode
Attention Mechanism — 4× in this paper · explained in 3 episodes
Embeddings — 4× in this paper · explained in 2 episodes
F1 Score — 4× in this paper · explained in 3 episodes
GNNs — 4× in this paper · explained in 2 episodes
Generalization — 4× in this paper · explained in 13 episodes
Interpretability — 4× in this paper · explained in 15 episodes
Late Fusion — 4× in this paper · explained in 2 episodes
Multimodal Fusion — 4× in this paper · explained in 3 episodes
Clustering — 3× in this paper · explained in 2 episodes
Multimodal Learning — 3× in this paper · explained in 2 episodes
Multimodal Reasoning — 3× in this paper · explained in 5 episodes
Robustness — 3× in this paper · explained in 5 episodes
Ablation Studies — 2× in this paper · explained in 1 episode
Abstraction — 2× in this paper · explained in 1 episode
Contrastive Learning — 2× in this paper · explained in 5 episodes
Diagnostic Tools — 2× in this paper · explained in 2 episodes
Diffusion Model — 2× in this paper · explained in 8 episodes
Diffusion Models — 2× in this paper · explained in 6 episodes
Adam — 1× in this paper · explained in 2 episodes
Feature Extraction — 1× in this paper · explained in 2 episodes
Federated Learning — 1× in this paper · explained in 15 episodes
Fine-Tuning — 1× in this paper · explained in 5 episodes
Graph Neural Networks (GNNs) — 1× in this paper · explained in 7 episodes
Parameter Efficiency — 1× in this paper · explained in 3 episodes
Semi-supervised Learning — 1× in this paper · explained in 4 episodes
Structured Reasoning — 1× in this paper · explained in 3 episodes
Trace — 1× in this paper · explained in 2 episodes

Summary

Automated skin cancer classification from dermoscopic images remains challenging due to heterogeneous lesion structure, strong intra-class variability, and subtle visual differences between benign and malignant cases. The gist: Explicitly modeling lesions as graphs of spatially coherent superpixel regions represented as frozen CNN features and introducing a dedicated metadata context node connected to all regions allows for structured integration of demographic/clinical variables within the same relational space.

Problem Addressed

Existing CNN/ViT pipelines typically rely on global or patch-level features and often combine patient metadata via late fusion, which limits spatially grounded multimodal reasoning. Many models operate on grid/patch tokens and do not provide an explicit, spatially grounded representation of lesion subregions and their geometric relations. Furthermore, metadata is commonly treated as an auxiliary vector rather than being integrated into the core reasoning process, reducing interpretability and limiting structured multimodal interactions. Graph neural networks (GNNs) offer a natural abstraction for region-centric reasoning by encoding regions in images as nodes and edges, but only a few dermoscopy graphs employ context-aware graph reasoning with edge-attributed geometry.

Methodology Overview

The GeoMeta-GT pipeline constructs a geometry-attributed superpixel graph and performs edge-aware attention-based message passing to classify lesions. The CNN backbone remains frozen, and all task-specific adaptation is confined to the graph module. The process involves several key steps:

Superpixel decomposition: Input images are decomposed into SLIC superpixels, forming region nodes where node descriptors are extracted by concatenating a global context statistic (mean-pooled) and a local saliency statistic (max-pooled) features over all pixels in the region.
Graph construction: A graph G = (V, E) is constructed with superpixel nodes VS = ⋃Rk, each associated with node feature DS = ⋃Dk, and a single metadata node Vmeta. Spatial edges ES are constructed by n-nearest neighbor search based on the Euclidean distance between centroid locations.
Metadata-as-node fusion: Patient metadata is embedded into the superpixel node feature space Dk using a learnable linear projection, and a dedicated metadata node Vmeta is introduced with feature vector Dmeta, connected to all superpixel nodes VS via edges Emeta = ⋃(Vmeta, k),(k, Vmeta).

Edge-aware Graph Transformer for Feature Enhancement

A novel edge-aware graph transformer performs relational reasoning by explicitly incorporating edge attributes into the attention mechanism. For a target node i, the attention score αi,j from node i (target) to node j (neighbor) is calculated by injecting the edge embedding into the key before the dotproduct: αi,j = softmax s Qi J(Kj + ˆei,j). Subsequently, the feature at node i is updated by aggregating messages from its neighbors: Dˆi = Qi + Σⱼ∈N(i) αi,j W4Dj + W5ei,j. This design allows the model to learn to emphasize edges that are most informative for the downstream task and enables edge attributes to shut down or boost the influence of specific neighbors, inducing locally adaptive region features.

Structural Refinement and Graph-level Readout

To encourage semantic consistency among connected superpixels, a lightweight similarity-weighted propagation step is applied to Dˆi. The refined embedding Zi is computed as Zi = Σj∈N(i)∪⋃γi,j Dˆj, where γi,j = softmax j∈N(i)∪⋃β cosine(Dˆi, Dˆj). This refinement step improves the discriminative patterns by aggregating information from neighboring nodes based on feature similarity. Finally, a graph-level representation is obtained via global mean pooling over the refined node embeddings, Z = MeanPool(Zi i∈V), which is then fed into a classifier for binary lesion classification.

Results and Contributions

Experiments on four public benchmarks demonstrate that explicit region-level relational modeling and graph-native multimodal fusion yield consistent gains over the state-of-the-art. The model consistently outperformed SOTA methods, achieving 98.61% accuracy on ISIC2024, 98.23% on HAM10000, 97.17% on PAD-UFES-20, and 95.41% on HIBA. Ablation studies confirm that incorporating geometric edge encoding (GEE) leads to consistent and substantial improvements across all datasets, with accuracy increasing from 96.14% to 98.61% on ISIC2024. Furthermore, the significance of metadata is confirmed as it substantially improves accuracy across all benchmarks by integrating patient metadata as a dedicated graph node. The backbone selection also impacts performance, with ResNet152 elevating accuracy to 98.

Improvements for AI systems

Here are the specific improvements to existing AI systems based on the GeoMeta-GT framework, and what these improved systems can achieve:

The GeoMeta-GT framework introduces a novel region-based graph learning architecture that explicitly models spatial relationships and integrates clinical metadata directly into the graph reasoning process. The primary improvements focus on moving from global or patch-level feature extraction to spatially grounded, multimodal relational modeling.

Here are the specific improvements and capabilities:

The system can perform highly accurate skin lesion classification (benign vs. malignant) by explicitly modeling the geometric arrangements of subregions within a lesion using a graph structure.
The improved AI system achieves state-of-the-art (SOTA) performance on challenging, heterogeneous public benchmarks like ISIC2024, HAM10000, PAD-UFES-20, and HIBA by leveraging geometry-aware edge encoding (distance and orientation).
The model can capture fine-grained lesion morphology—such as irregular borders and varied pigmentation patterns—which are critical for distinguishing subtle malignant cases from benign ones.
It enables structured multimodal reasoning by integrating patient metadata (age, sex, clinical context) not via late fusion (concatenation), but through a dedicated metadata context node connected to all lesion regions within the graph structure.
The architecture utilizes an edge-aware Graph Transformer that performs locally adaptive aggregation, meaning it can selectively emphasize informative neighboring regions while suppressing noise from irrelevant ones based on geometric relevance.
The system gains robustness against imaging noise and acquisition variability (as shown by superior performance on the HIBA dataset) because the graph structure provides a more stable representation than purely global or sequential attention mechanisms.
The model is parameter-efficient because it relies on a frozen CNN backbone, confining task-specific learning only to the graph module, ensuring consistent performance across different datasets without requiring massive retraining of deep feature extractors.

This improved AI system can perform the following specific tasks:

Identify and classify skin lesions (e.g., melanoma) with significantly higher accuracy (up to 98%+ on ISIC2024) compared to existing CNN/ViT pipelines that rely solely on global features or simple patch tokens.
Provide a spatially grounded justification for its classification by analyzing the geometric relationships between different subregions of the lesion (e.g., The malignant classification is driven by the irregular distance and orientation between Region A and Region B).
Improve diagnostic confidence in real-world clinical settings by providing a unified representation that simultaneously encodes both visual pathology (via superpixel features) and relevant patient clinical context (via metadata).
Serve as a robust front-end for triage systems where the model can handle diverse lesion appearances and imaging conditions without significant performance degradation.