A new preprocessing approach to improve the performance of CNN-based skin lesion classification

Area of Science:

• Dermatology
• Medical Imaging
• Artificial Intelligence

Background:

• Skin lesions, including melanoma, pose a global health risk, necessitating early detection for improved patient outcomes.
• Accurate diagnosis is challenging due to visual similarities between lesion types, low contrast, and image artifacts.
• Convolutional neural networks (CNNs) show promise for automated skin lesion classification.

Purpose of the Study:

• To develop and evaluate a novel preprocessing technique for skin lesion image analysis.
• To assess the impact of region of interest (RoI) extraction on CNN model performance.
• To compare the accuracy and efficiency of CNNs trained on raw versus preprocessed datasets.

Main Methods:

• A new preprocessing method was developed to extract the region of interest (RoI) from skin lesion images.
• State-of-the-art CNN classifiers were trained and evaluated using both raw and RoI-extracted image datasets.
• Performance metrics included prediction accuracy, training time, and evaluation (inference) time.

Main Results:

• Training CNN models with RoI-extracted datasets significantly improved prediction accuracy (e.g., InceptionResNetV2 showed a 2.18% increase).
• The RoI extraction preprocessing technique substantially decreased both training and evaluation times for the classifiers.
• The proposed method demonstrated enhanced efficiency and effectiveness in automated skin lesion analysis.

Conclusions:

• The developed RoI extraction preprocessing technique is effective in enhancing CNN-based skin lesion classification.
• This approach offers a valuable tool for improving the accuracy and efficiency of early skin cancer detection.
• Optimized CNN models using RoI-extracted data contribute to faster and more reliable diagnostic processes.