Domain-specific classification-pretrained fully convolutional network encoders for skin lesion segmentation.

Area of Science:

• Medical image analysis
• Deep learning for dermatology
• Computational pathology

Background:

• Automated skin lesion segmentation is crucial for dermatological diagnosis.
• Fully convolutional neural networks (FCNNs) show promise for this task.
• Investigating transfer learning strategies can optimize FCNN performance.

Purpose of the Study:

• To evaluate the impact of transferring pre-trained encoder weights on skin lesion segmentation accuracy.
• To compare domain-specific pretraining versus general ImageNet pretraining for segmentation tasks.
• To assess the influence of ground truth annotation complexity on segmentation performance.

Main Methods:

• A U-Net style architecture was employed, reusing ResNet34 layers as encoders.
• Encoding layers were initialized randomly, with ImageNet weights, or with weights fine-tuned on skin lesion classification.
• The network was trained for binary segmentation on the ISIC 2017 dataset.

Main Results:

• Pretraining with ImageNet or fine-tuned weights yielded higher Jaccard indices (0.763, 0.768) than random initialization (0.740).
• Conditional random fields post-processing decreased performance on simple shapes but improved it for complex annotations.
• Annotation complexity significantly impacts segmentation metrics.

Conclusions:

• Domain-specific pretraining of encoders is beneficial for skin lesion segmentation, particularly with limited ground truth data.
• ImageNet pretraining may suffice when ample segmentation data is available.
• Annotation complexity is a critical factor in skin lesion segmentation research.