A shallow deep learning approach to classify skin cancer using down-scaling method to minimize time and space...

Area of Science:

• Dermatology
• Medical Imaging
• Computer Science

Background:

• Automated melanoma recognition is challenging due to complex lesion features, low contrast, inter-class similarity, and artifacts.
• Existing computer-aided solutions for skin cancer classification face limitations.

Purpose of the Study:

• To propose a shallow deep learning model for classifying skin lesions as benign or malignant.
• To address challenges of limited training data, overfitting, and computational complexity in medical image analysis.

Main Methods:

• A shallow convolutional neural network (SCNN_12) was developed.
• Image preprocessing and data augmentation techniques were applied to a dataset of 16,485 images.
• The 'box blur' down-scaling method was used to enhance efficiency.

Main Results:

• The SCNN_12 model achieved an accuracy of 98.87% using the Adam optimizer and a 0.001 learning rate.
• Ablation studies determined optimal model parameters and hyperparameters.
• K-fold cross-validation and varied dataset splits confirmed no overfitting; robustness was tested on noisy data.

Conclusions:

• Effective training for medical image analysis is achievable with lightweight networks and limited data.
• The proposed SCNN_12 model offers an efficient solution for automated melanoma recognition, balancing accuracy and computational resources.