FFTL-Net: a network for the classification of skin diseases based on feature fusion and transfer learning
The objective is to address the issues of data imbalance, overfitting, and inadequate generalization ability in skin disease datasets and recognition models. The proposed model for the classification of skin diseases is based on the fusion of features and the utilization of transfer learning. The mo...
Saved in:
| Main Authors: | , , , , , , , |
|---|---|
| Format: | Article |
| Language: | English |
| Published: |
IOP Publishing
2025-01-01
|
| Series: | Machine Learning: Science and Technology |
| Subjects: | |
| Online Access: | https://doi.org/10.1088/2632-2153/ade4f0 |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | The objective is to address the issues of data imbalance, overfitting, and inadequate generalization ability in skin disease datasets and recognition models. The proposed model for the classification of skin diseases is based on the fusion of features and the utilization of transfer learning. The model’s architecture is predicated on a dense connection network that serves as its fundamental framework, with LBP and HOG features incorporated as supplementary inputs. Subsequently, a feature fusion module integrated with an attention mechanism is employed to extract and combine features. Finally, the Softmax-loss function of category equilibrium and the domain adaptive strategy based on the maximum mean difference are established. The integration of prior knowledge into the deep network is a critical step in addressing the challenges of overfitting and data imbalance in skin disease classification. The FFTL-Net achieved AUC value of 98.16% on the International skin imaging collaboration (ISIC) 2018 dataset and 98.31% on the ISIC 2019 dataset. This represents an improvement of 1.25% and 0.33% compared to the second-ranked algorithm, respectively. The experimental results demonstrate the efficacy of the model in addressing the data imbalance issue in skin disease datasets, with prediction accuracies of at least 93% being achieved for BCC and other rare samples. The model demonstrates superior recognition accuracy, augmented generalization capability, and an absence of indications of overfitting. |
|---|---|
| ISSN: | 2632-2153 |