💻 computer science

WaveInvNet: A Wavelet-Separated Dynamic Involution Network for Multimodal Skin Lesion Classification

This paper proposes WaveInvNet, a lightweight multimodal deep learning architecture that integrates dermoscopic images with patient metadata using a novel Wavelet-Separated Dynamic Involution block to adaptively capture heterogeneous lesion features, achieving state-of-the-art performance in skin lesion classification on the HAM10000 dataset.

Original authors: Mohamed Amine Ibrahimi, Abdelghafour Abbad, Rachid Ben Abbou, Khalid Abbad

Published 2026-09-23

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Mohamed Amine Ibrahimi, Abdelghafour Abbad, Rachid Ben Abbou, Khalid Abbad

Original paper licensed under CC BY 4.0 (https://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

Skin cancer is a relentless global challenge, with millions of new cases diagnosed every year. Among these, melanoma is particularly dangerous because it can spread rapidly to other parts of the body, making early detection a matter of life and death. For decades, doctors have relied on dermoscopy, a technique that uses a specialized magnifying tool to examine skin lesions in detail. However, this process depends heavily on the experience of the clinician, and even experts can disagree on what they see. To help, scientists have turned to artificial intelligence, teaching computers to recognize the subtle patterns of disease in medical images. While these computer systems have become quite good at spotting abnormalities, they often struggle with the complexity of skin lesions. Traditional systems treat every part of an image the same way, using a single, unchanging set of rules to scan the entire picture. This approach can miss the unique textures of a lesion's center or the irregular edges of its border, failing to adapt to the specific details that distinguish a harmless mole from a dangerous tumor.

A team of researchers from Sidi Mohamed Ben Abdellah University in Morocco has developed a new approach to solve this problem, creating a system they call WaveInvNet. Instead of using a one-size-fits-all method, their system is designed to look at a skin image in two different ways simultaneously: one that captures the broad, overall shape of the lesion, and another that focuses on the fine, intricate details. They achieved this by using a mathematical technique known as a wavelet transform, which acts like a prism for images, splitting visual information into separate layers of structure and texture. By separating these layers, the computer can analyze the big picture and the tiny details independently, ensuring that no critical diagnostic clue is lost in the process.

The researchers built their system around a novel component they named a Wavelet-Separated Dynamic Involution block. In simpler terms, this is a smart processing unit that learns to adjust its own focus based on what it sees. When the system encounters a smooth area of skin, it applies one set of analytical rules; when it encounters a rough, irregular border, it instantly switches to a different set of rules that are better suited for that specific texture. This dynamic adjustment allows the computer to be much more sensitive to the unique characteristics of each lesion. Furthermore, the system does not rely on the image alone. It also takes into account patient information, such as age, sex, and the location of the lesion on the body. By combining the visual analysis with these clinical facts, the system creates a more complete picture of the patient's condition before making a diagnosis.

To test their creation, the researchers trained the system using a large collection of over 10,000 skin images from the HAM10000 dataset, which includes seven different types of skin lesions. They split the data so that the system learned from most of the images and was then tested on the ones it had never seen before. The results were striking. The new system correctly identified the type of skin lesion in 93.63% of the test cases, a performance that surpassed several other well-known artificial intelligence models. It also achieved a score of 98.46% in its ability to distinguish between different classes of disease, a metric known as the area under the curve. Perhaps most importantly, the system accomplished this high level of accuracy while using significantly fewer computer resources than its competitors. It required only 2.54 million adjustable settings, which is less than half the number needed by some of the other leading models, making it a lightweight and efficient tool for medical use.

The study also revealed where the system excelled and where it still faced challenges. It was perfect at identifying certain rare types of lesions, such as vascular lesions and dermatofibromas, with zero errors in the test set. However, it did occasionally confuse melanoma with benign moles, which are the two most common and visually similar categories. This confusion is a known difficulty in the field, as the visual differences between a dangerous melanoma and a harmless mole can be incredibly subtle. Despite these minor errors, the system demonstrated a consistent ability to separate the different classes of disease, proving that its method of splitting the image into structural and detailed components works effectively. The researchers noted that while the system performed well on this specific dataset, future work would need to test it on a wider variety of images and clinical conditions to ensure it works reliably in real-world hospitals.

This work represents a significant step forward in how artificial intelligence can assist doctors in diagnosing skin cancer. By moving away from rigid, uniform analysis and embracing a method that adapts to the specific details of each image, the researchers have shown that machines can learn to see skin lesions more like human experts do. The integration of patient data with visual analysis further strengthens the diagnosis, mirroring the way a doctor considers the whole patient, not just the spot on their skin. As the burden of skin cancer continues to rise, tools like WaveInvNet offer a promising path toward faster, more accurate, and more accessible care, potentially saving lives by catching disease earlier than ever before.

Technical Summary: WaveInvNet

Problem Statement
The paper addresses the critical need for early and accurate skin cancer diagnosis to reduce mortality, particularly for melanoma. While deep learning, specifically Convolutional Neural Networks (CNNs), has advanced automated dermoscopic image analysis, conventional architectures face limitations. Standard CNNs rely on spatially shared kernels that process all image regions uniformly, failing to adapt to the heterogeneous structures of skin lesions (e.g., varying centers, boundaries, and surrounding skin). Furthermore, conventional feature extraction often processes structural information and fine-grained diagnostic details jointly, risking the loss of subtle patterns during downsampling. Additionally, skin lesion datasets like HAM10000 exhibit significant class imbalance, which can bias models toward majority classes.

Methodology
The authors propose WaveInvNet, a lightweight multimodal architecture designed to integrate dermoscopic images with patient metadata (age, sex, and anatomical location). The core innovation lies in the Wavelet-Separated Dynamic Involution (WSDI) block, organized within a four-stage hierarchical encoder.

Wavelet-Separated Dynamic Involution (WSDI) Block:
- Stationary Haar Wavelet Decomposition: Input feature maps undergo a stationary Haar wavelet transform. This separates the data into a low-frequency structural component ( $F_{LL}$ ) and three high-frequency detail components ( $F_{LH}, F_{HL}, F_{HH}$ ). Unlike discrete wavelet transforms, this method preserves the original spatial resolution without downsampling.
- Frequency-Specific Kernel Generation: The block employs separate generators for structure, detail, and gating.
  - The structure branch uses the low-frequency subband to generate raw kernel logits, which are normalized via Softmax to create a non-negative, unit-sum kernel ( $K_{str}$ ) for structural aggregation.
  - The detail branch processes the concatenated high-frequency subbands. The resulting logits are bounded by a hyperbolic tangent, mean-centered, and L1-normalized to produce a signed, zero-sum kernel ( $K_{det}$ ) that emphasizes local variations.
  - The gating branch uses absolute high-frequency coefficients to generate a spatially varying gate ( $\beta$ ) that modulates the contribution of the detail kernel.
- Adaptive Kernel Integration: The final kernel is computed as $K_{wsdi} = K_{str} + \beta \odot K_{det}$ . This allows the network to dynamically adjust how much fine-grained detail influences the aggregation at each spatial location.
- Grouped Involution: These adaptive kernels are applied to local neighborhoods using a grouped involution operator, where kernels are shared across channel groups but vary by spatial location.
- Residual Refinement: The output undergoes channel expansion, GELU activation, projection, LayerScale, and stochastic depth before being added to the residual input.
Multimodal Fusion:
- Patient metadata is standardized and encoded.
- After the four-stage visual encoder extracts frequency-aware features, global average pooling is applied.
- The resulting visual representation is concatenated with the encoded metadata before passing through a dense layer with softmax activation for final classification.
Data Handling:
- The model was trained on the HAM10000 dataset (10,015 images, 7 classes).
- To address class imbalance, targeted data augmentation was applied to underrepresented classes (e.g., augmenting VASC and DF samples significantly).
- Missing age values were imputed with the median.

Key Contributions

Novel WSDI Block: The paper introduces the first application of a mechanism that uses wavelet-derived structural and detail components to directly guide the generation of location-specific involution kernels for skin lesion classification.
Lightweight Architecture: WaveInvNet is a four-stage hierarchical encoder that progressively learns frequency-aware and spatially adaptive representations with a low parameter count.
Comprehensive Evaluation: The study provides a unified comparison against established CNN architectures (EfficientNetB0, DenseNet121, ResNet50, ConvNeXt-Base, InceptionV3) and recent state-of-the-art approaches, evaluating performance, computational efficiency, and model complexity.

Results
Evaluated on the HAM10000 test set (70/10/20 split, averaged over 5 runs), WaveInvNet achieved:

Accuracy: 93.63%
AUC: 98.46%
F1-Score: 94.81%
Precision: 94.45%
Recall: 95.20%

The model demonstrated consistent performance with low standard deviations across runs. In terms of efficiency, WaveInvNet required only 2.54 million trainable parameters (10.16 MB for weights), representing a ~42.6% reduction compared to the smallest baseline (EfficientNetB0) and an ~88.7% reduction compared to InceptionV3.

Confusion matrix analysis revealed that the model classified AKIEC, DF, and VASC without error. The primary confusion occurred between Melanoma (MEL) and Nevus (NV), and between Benign Keratosis-like lesions (BKL) and NV, which is consistent with known diagnostic challenges in dermatology.

Significance and Claims
The paper claims that WaveInvNet demonstrates that combining frequency-aware visual processing (via wavelet separation) with spatially adaptive aggregation (via dynamic involution) and patient metadata enables accurate, parameter-efficient multiclass skin-lesion classification. The authors assert that their approach outperforms existing deep learning baselines under identical experimental conditions while maintaining a significantly smaller model footprint. They position the work as a step toward more adaptive and efficient diagnostic support systems that can handle the heterogeneity of skin lesions and the imbalance of clinical datasets.

The authors note that while results are promising, future work should include testing on additional datasets to verify generalization across different clinical conditions, exploring stratified k-fold cross-validation, and extending the metadata branch to include other clinical variables. They also identify explainability and uncertainty estimation as areas for future investigation to enhance clinical interpretability.

More like this