PaperlayerView exact revision on arXiv

Source-linked AI summary

Skin Lesion Analysis Towards Melanoma Detection Using Deep Learning Network

Yuexiang Li, Linlin Shen

arXiv:1703.00577v2cs.CV

TL;DR

The paper targets difficult automatic melanoma analysis in dermoscopy images across the three ISIC 2017 tasks. It introduces LIN for joint segmentation and classification, LICU for refining lesion possibilities, and LFN for dermoscopic feature extraction. On the ISIC 2017 testing set, the frameworks achieved reported task scores of 0.718, 0.833, and 0.823.

  • Problem

    Automatic melanoma recognition is difficult because dermoscopy images can have low lesion–skin contrast and visually similar melanoma and non-melanoma lesions, while dermoscopic feature extraction remains a new task with few prior studies.

  • Method

    LIN uses two fully-convolutional residual networks and a lesion index calculation unit for lesion segmentation and classification, while LFN uses a CNN for dermoscopic feature extraction.

  • Results

    The frameworks achieved JA 0.718 for lesion segmentation, AUC 0.823 for lesion classification, and LFN average precision 0.409 plus sensitivity 0.665 for dermoscopic feature extraction.

  • Takeaways & Limitations

    The paper provides deep learning frameworks addressing all three ISIC 2017 skin-lesion image-processing tasks, with LFN offering a CNN-based approach for the relatively new feature-extraction task.

Abstract

from arXiv · show

Skin lesion is a severe disease in world-wide extent. Early detection of melanoma in dermoscopy images significantly increases the survival rate. However, the accurate recognition of melanoma is extremely challenging due to the following reasons, e.g. low contrast between lesions and skin, visual similarity between melanoma and non-melanoma lesions, etc. Hence, reliable automatic detection of skin tumors is very useful to increase the accuracy and efficiency of pathologists. International Skin Imaging Collaboration (ISIC) is a challenge focusing on the automatic analysis of skin lesion. In this paper, we proposed two deep learning methods to address all the three tasks announced in ISIC 2017, i.e. lesion segmentation (task 1), lesion dermoscopic feature extraction (task 2) and lesion classification (task 3). A deep learning framework consisting of two fully-convolutional residual networks (FCRN) is proposed to simultaneously produce the segmentation result and the coarse classification result. A lesion index calculation unit (LICU) is developed to refine the coarse classification results by calculating the distance heat-map. A straight-forward CNN is proposed for the dermoscopic feature extraction task. To our best knowledges, we are not aware of any previous work proposed for this task. The proposed deep learning frameworks were evaluated on the ISIC 2017 testing set. Experimental results show the promising accuracies of our frameworks, i.e. 0.718 for task 1, 0.833 for task 2 and 0.823 for task 3 were achieved.

I. INTRODUCTION

Melanoma recognition is clinically important but difficult because dermoscopy images contain low lesion–skin contrast, visually similar lesion types, and varied skin conditions. The paper addresses all three ISIC 2017 tasks with deep learning frameworks, including a combined segmentation/classification network and a CNN for dermoscopic features.

  • Automatic melanoma recognition remains difficult because lesions may have low contrast with skin, resemble non-melanoma lesions, and occur under varied skin conditions.
  • ISIC 2017 introduced lesion segmentation, dermoscopic feature extraction, and lesion classification as three image-processing tasks.
  • LIN uses multi-scale fully-convolutional residual networks and a lesion index calculation unit to jointly address lesion segmentation and classification.
  • LFN is a CNN-based framework for dermoscopic feature extraction, a relatively new ISIC task with few prior studies.

II. METHODS

This section introduces the deep learning methods developed for the paper’s different skin-lesion analysis tasks.

  • The paper develops deep learning methods for different skin-lesion image-processing tasks.

A. Lesion segmentation and classification (task 1 & 3)

The segmentation and classification pipeline crops and proportionally resizes images, trains separate augmented-data FCRNs, and uses LICU to refine coarse lesion possibilities with distance-based pixel importance.

  • Pre-processing: Center cropping followed by proportional downscaling reduces computation while enlarging the lesion area and preserving its shape.
  • Data Augmentation: Class-balanced augmentation rotates images by category-specific angles, then random x- or y-axis flips produce a second dataset for FCRN training.
  • Lesion Indexing Network (LIN): LIN extends FCRN-88 with two FCRNs trained on differently augmented datasets to jointly perform lesion segmentation and classification.
  • Lesion Indexing Network (LIN): Separate FCRNs outperform one FCRN-88 trained on combined DR+DM data because the mirror operation can fool the network and encourage overfitting.
  • Lesion Index Calculation Unit (LICU): LICU refines summed coarse possibility maps by normalizing them and weighting lesion pixels according to distance from the nearest border.

B. Dermoscopic feature extraction (task 2)

The Lesion Feature Network addresses ISIC 2017’s dermoscopic feature extraction task by classifying augmented, uniformly sized superpixel patches with a CNN.

  • The new ISIC 2017 dermoscopic feature extraction task aims to extract clinical features from dermoscopic images.
  • Superpixels extraction: Dermoscopic images are subdivided into superpixels, whose contents are resized to 56x56 patches for the Lesion Feature Network.The superpixel masks contain Pigment Network, Negative Network, Streaks, Milia-like Cysts, and background categories.
  • Data augmentation: Data augmentation balances the extremely imbalanced patch dataset using random sampling and patch rotation.Patch rotation increases volumes to 12,908 for NN, 18,424 for MC, and 8,324 for S.
  • Lesion Feature Network: LFN uses 12 convolutional layers arranged in four stages, with network-in-network structures, max and average pooling, and softmax loss.Each stage contains three convolutional layers, and 1x1 convolutions integrate features from 3x3 convolutions.
  • Lesion Feature Network: Class weighting addresses residual imbalance by assigning weights 1, 1, 5, 3, and 8 to B, PN, NN, MC, and S, respectively.
  • Implementation: The network uses an 80:20 training-validation split, SGD with learning rate 0.01 and momentum 0.9, and converged after 10 epochs on a GeForce GTX TITAN X.

A. Datasets

ISIC 2017 provides annotated training, validation, and testing datasets supporting lesion segmentation, dermoscopic feature extraction, and classification evaluation.

  • ISIC 2017 provides 2,000 training images with segmentation masks, superpixel masks, and classification annotations.The classification categories are Melanoma, Seborrheic keratosis, and Nevus.
  • The dataset examples pair original lesion images with segmentation masks and superpixel masks for dermoscopic feature extraction.
  • ISIC 2017 includes a validation set of 150 images for framework evaluation.
  • The final competition uses a testing set of 600 skin lesion images, assessed through an online submission system because ground truths were initially unpublished.The organizers later released validation and testing ground truths.
  • The paper analyzes LIN and LFN on the validation set and reserves testing-set state-of-the-art comparisons for the next section.

B. Evaluation Metrics

The challenge evaluates segmentation with overlap and classification-oriented measures, while AUC provides the primary ranking metric for dermoscopic feature extraction and lesion classification.

  • Lesion segmentation: Lesion segmentation is evaluated using accuracy, Jaccard Index, Dice coefficient, sensitivity, and specificity.The metrics are defined from true-positive, true-negative, false-positive, and false-negative counts.
  • Lesion segmentation: Segmentation submissions are ranked by the Jaccard Index, while the other metrics serve as reference measures.
  • Dermoscopic feature extraction & Lesion classification: Dermoscopic feature extraction and lesion classification use accuracy, sensitivity, specificity, average precision, and AUC.AUC is generated by evaluating sensitivity against false-positive rate across threshold settings.

C. Lesion Indexing Network (LIN)

LIN combines lesion segmentation and classification, using multi-scale inputs and LICU refinement to address difficult dermoscopy images. Validation experiments show gains from separate augmented-data networks, multi-scale processing, and LICU.

  • Performance on lesion segmentation: Segmentation examples compare LIN outlines with ground truths across melanoma, seborrheic keratosis, and nevus images.The examples include low lesion-to-skin contrast and hair near lesion regions.
  • Training with DR and DM: LIN uses FCRN-88 networks trained with separate DR and DM datasets; combining them causes overfitting and worse validation performance.The combined DR+DM model has the lowest JA, 0.607, whereas LIN achieves 0.710 on the validation set.
  • Multi-scale input images: 0.753 JA is achieved by LIN using all three input scales, exceeding the second-ranked two-scale configuration by 0.002.The two-scale ~300 and ~500 configuration is preferred for applications because of computational efficiency.
  • Performance of LICU: LICU improves LIN’s classification result by 0.017 over plain LIN, reaching 0.908 AUC on the validation set.LICU estimates pixel importance with a distance map and refines FCRN-88 possibility maps.

D. Lesion Feature Network (LFN)

LFN is evaluated through architecture, loss, and normalization experiments for dermoscopic feature extraction. Its performance depends on sufficient layer width, weighted softmax loss, and batch normalization.

  • Network architecture: 0.848 AUC is achieved by the original LFN, exceeding Narrow LFN at 0.822 and Wide LFN at 0.803.Narrow layers reduce feature-representation capacity, while overly wide layers lead to overfitting.
  • Weighted softmax loss: 0.778 AUC results without weighted softmax loss, demonstrating the importance of WSL for the imbalanced training set.Data augmentation alone does not fully balance category sample volumes.
  • Usage of Batch Normalization: Batch normalization improves LFN AUC by 0.006.BN is placed between convolutional layers and rectified linear units layers.

IV. COMPARISON WITH STATE-OF-THE-ART

The frameworks were compared with other submissions on the ISIC 2017 testing set across segmentation, dermoscopic feature extraction, and classification. LIN ranked ninth in task 1, second in task 2, and tied for twelfth in task 3.

  • Lesion segmentation: 0.718 JA placed the method ninth among 21 lesion-segmentation submissions in ISIC 2017.The final ranking for task 1 is based on JA.
  • Dermoscopic feature extraction: 0.833 AUC placed the method second among three dermoscopic-feature-extraction submissions.AUC is the final ranking metric for task 2.
  • Lesion classification: 0.823 AUC produced a tie for twelfth place among 23 lesion-classification submissions.AUC is the final ranking metric for task 3.

V. CONCLUSION

The paper proposes LIN and LFN for the three ISIC 2017 skin-lesion processing tasks. LIN jointly handles segmentation and classification, while LFN addresses dermoscopic feature extraction and achieves strong reported testing-set results.

  • V. CONCLUSION: LIN and LFN address lesion segmentation, dermoscopic feature extraction, and lesion classification.LIN covers segmentation and classification; LFN covers dermoscopic feature extraction.
  • V. CONCLUSION: LIN uses fully-convolutional residual networks with LICU to produce segmentation and refined classification results.LICU measures pixel importance with a distance map and refines the coarse classification result.
  • V. CONCLUSION: LFN is a CNN trained on patches extracted from superpixel masks for dermoscopic feature extraction.The paper states that it is unaware of previous work for this task.
  • V. CONCLUSION: 0.718 JA for segmentation and 0.823 AUC for classification were comparable to competition winners, while LFN achieved best average precision and sensitivity of 0.409 and 0.665.These results were obtained on the ISIC 2017 testing set.
Loading 1703.00577v2…