Source-linked AI summary
Automatic Skin Lesion Analysis using Large-scale Dermoscopy Images and Deep Residual Networks
Lei Bi, Jinman Kim, Euijoon Ahn, Dagan Feng
TL;DR
Automated skin lesion analysis aims to distinguish clinically relevant lesion types despite the challenges of dermoscopy images. This paper uses deep residual networks for lesion segmentation and classification, with its proposed multi-scale approach achieving first place on the validation set.
Problem
Skin lesion analysis must distinguish melanoma, seborrheic keratosis, and nevus as separate lesion categories.
Method
The paper applies deep residual networks to learn visual representations for lesion segmentation and classification, including a multi-scale segmentation approach.
Results
The proposed multi-scale approach achieved 1st place on the validation set by the submission deadline.
Takeaways & Limitations
The proposed multi-scale approach was highly competitive in the ISIC 2017 lesion segmentation challenge.
Abstract
from arXiv · showhide
Malignant melanoma has one of the most rapidly increasing incidences in the world and has a considerable mortality rate. Early diagnosis is particularly important since melanoma can be cured with prompt excision. Dermoscopy images play an important role in the non-invasive early detection of melanoma [1]. However, melanoma detection using human vision alone can be subjective, inaccurate and poorly reproducible even among experienced dermatologists. This is attributed to the challenges in interpreting images with diverse characteristics including lesions of varying sizes and shapes, lesions that may have fuzzy boundaries, different skin colors and the presence of hair [2]. Therefore, the automatic analysis of dermoscopy images is a valuable aid for clinical decision making and for image-based diagnosis to identify diseases such as melanoma [1-4]. Deep residual networks (ResNets) has achieved state-of-the-art results in image classification and detection related problems [5-8]. In this ISIC 2017 skin lesion analysis challenge [9], we propose to exploit the deep ResNets for robust visual features learning and representations.
1 Introduction · 2 Methods
The paper applies deep residual networks to automated dermoscopy analysis in the ISIC 2017 challenge, targeting robust visual feature learning and lesion classification. Its methods combine FCN-based lesion segmentation with multi-class, binary, and ensemble classification strategies.
- 2.1 Deep Residual Networks: ResNets are exploited for robust visual feature learning and representations in the ISIC 2017 skin lesion analysis challenge.The paper motivates ResNets by their state-of-the-art performance in image classification and detection.
- 2.2 Lesion Segmentation: Lesion segmentation uses a fully convolutional network architecture with convolutional and deconvolutional layers to upsample ResNet feature maps into score masks.Segmentation is presented as a fundamental requirement for automated skin lesion analysis.
- 2.2 Lesion Segmentation: ~9,800 images were used for segmentation training and ~200 for validation, combining 2,000 challenge images with ~8,000 expert-annotated ISIC archive images.Images were downscaled to 500 pixels along the longer axis.
- 2.3 Lesion Classification: Classification labels lesions as melanoma, seborrheic keratosis, or nevus, distinguishing malignant from benign tumors and melanocytic from non-melanocytic origins.Melanoma and nevus are melanocytic, while seborrheic keratosis is non-melanocytic.
- 2.3 Lesion Classification: The classification methods include a 3-label multi-class model, two binary classifiers, and an ensemble combining the first two approaches.The binary classifiers separate melanoma from other lesions and seborrheic keratosis from other lesions.
- 2.3 Lesion Classification: ~3,600 images were used for classification training, including 2,000 challenge images and ~1,600 histology- or expert-confirmed ISIC archive images.ResNet’s final layer was modified to contain 3 or 2 neurons, and images were downscaled to 224 pixels along the shorter axis.
- 2.3 Lesion Classification: Classification training used 150 epochs, a linear learning-rate schedule with base learning rate 0.01, batch size 90, and random crops and flips.Each ResNet trained for around half a day on two Titan X GPUs.
3 Experiments and Results
Experiments evaluated segmentation and classification methods on a 150-image validation set from the challenge organizer. The proposed multi-scale segmentation approach achieved first place on the validation set, while classification approaches were assessed using averaged AUC.
- Evaluation setup: The evaluation used the challenge organizer’s validation set containing 150 skin lesion images, with results processed through the online submission system.Ground-truth labels were not publicly available.
- Segmentation: ResNet-Seg was compared with a VGGNet-based FCN, including experiments with additional training images and the proposed multi-scale integration approach.These experiments evaluated segmentation performance across the proposed method, baseline architecture, training data, and integration strategy.
- Segmentation: The proposed multi-scale approach achieved 1st place on the validation set by the submission deadline.The ranking was based on the challenge submission results.
- Classification: Three classification approaches—multi-class, binary, and ensemble—were evaluated on the 150 validation images using averaged area under curve (AUC).The evaluation covered the proposed classification variants under a common AUC-based criterion.