Table 2 Dataset details before and after data pre-processing.
From: Advancing skin cancer diagnosis with deep learning and attention mechanisms
Class | Original image count | Augmented image count | Total image count | Class imbalance | Handling imbalance & overfitting |
|---|---|---|---|---|---|
Actinic Keratosis | 327 | 654 | 981 | Low (underrepresented) | Augmented via flipping and rotation to balance the class |
Basal Cell Carcinoma | 514 | 1028 | 1542 | Moderate | Augmentation and model regularization are applied to prevent overfitting |
Dermatofibroma | 115 | 230 | 345 | Very Low | Increased augmentation and balanced sampling during training |
Melanoma | 1113 | 2226 | 3339 | Moderate (overrepresented) | An augmented, weighted loss function is used to prevent overfitting |
Nevus | 1707 | 3414 | 5121 | High (overrepresented) | Controlled sampling in mini-batches and augmentation techniques |
Pigmented Benign Keratosis | 1615 | 3230 | 4845 | High (overrepresented) | Data augmentation and regularization are applied to reduce the overfitting risk. |
Seborrheic Keratosis | 1232 | 2464 | 3696 | Moderate | Augmentation and class balancing strategies during model training |
Squamous Cell Carcinoma | 327 | 654 | 981 | Low (underrepresented) | Balanced via augmentation and targeted sampling |
Vascular Lesion | 142 | 284 | 426 | Very Low | Significant augmentation to balance class distribution |
Total | 8640 | 17,280 | 25,920 | - | Balancing strategies used for both underrepresented and overrepresented classes |
