ChimeraNet: U-Net for Hair Detection in Dermoscopic Skin Lesion Images

Abstract

Hair and ruler mark structures in dermoscopic images are an obstacle preventing accurate image segmentation and detection of critical network features. Recognition and removal of hairs from images can be challenging, especially for hairs that are thin, overlapping, faded, or of similar color as skin or overlaid on a textured lesion. This paper proposes a novel deep learning (DL) technique to detect hair and ruler marks in skin lesion images. Our proposed ChimeraNet is an encoder-decoder architecture that employs pretrained EfficientNet in the encoder and squeeze-and-excitation residual (SERes) structures in the decoder. We applied this approach at multiple image sizes and evaluated it using the publicly available HAM10000 (ISIC2018 Task 3) skin lesion dataset. Our test results show that the largest image size (448 × 448) gave the highest accuracy of 98.23 and Jaccard index of 0.65 on the HAM10000 (ISIC 2018 Task 3) skin lesion dataset, exhibiting better performance than for two well-known deep learning approaches, U-Net and ResUNet-a. We found the Dice loss function to give the best results for all measures. Further evaluated on 25 additional test images, the technique yields state-of-the-art accuracy compared to 8 previously reported classical techniques. We conclude that the proposed ChimeraNet architecture may enable improved detection of fine image structures. Further application of DL techniques to detect dermoscopy structures is warranted.

Keywords: Deep learning; Dermoscopy; Hair removal; Image segmentation; Melanoma; Transfer learning.

Fig. 1

Skin lesion dermoscopy images from the HAM10000 (ISIC2018 Task 3) dataset showing dark hair, light hair, and ruler marks

Fig. 2

Manually drawn hair masks corresponding to skin lesion images shown in Fig. 1. The hair masks include dark hair, white hair, and ruler marks

Fig. 3

Proposed ChimeraNet architecture for hair segmentation. An encoder-decoder architecture with pretrained EfficientNet model as the encoder network and the decoder network comprised of five squeeze-and-excitation residual blocks. Five skip-connections (red color) from the encoder to the decoder

Fig. 4

Double convolution block (left) and SERes block (right)

Fig. 5

Hair segmentation results of proposed ChimeraNet against U-Net and ResUNet-a for HAM10000 test images. The segmentation results show true positives (green), false positives (red), and false negatives (blue). The U-Net model finds more false positives (for example, gel bubbles), and ResUNet-a finds less hair. The proposed ChimeraNet model successfully detects hair with fewer false positives and false negatives

Fig. 6

Lesion image, ground-truth hair mask, and overlays of predicted hair mask on lesion image for nine hair detection methods. The proposed ChimeraNet method accurately detects more hair with less noise compared to other hair detection methods