Source-linked AI summary
Improving Dermoscopic Image Segmentation with Enhanced Convolutional-Deconvolutional Networks
Yading Yuan, Yeh-Chi Lo
TL;DR
Automatic skin-lesion segmentation is important for melanoma computer-aided diagnosis but is challenged by lesion variation and large image datasets. The paper extends CDNN with deeper small-kernel architecture and multiple color-space inputs, achieving a 0.765 average Jaccard Index on 600 ISBI 2017 testing images and ranking first in the challenge.
Problem
Automatic skin-lesion segmentation is essential for melanoma computer-aided diagnosis but is difficult because lesion appearances vary substantially across patients and datasets can be large.
Method
The method extends CDNN with a deeper small-kernel architecture and adds HSV and CIELAB channels to RGB inputs, using a Jaccard-distance-based training loss.
Results
0.765 average Jaccard Index (JA) was achieved on 600 ISBI 2017 challenge testing images using 2000 training images, ranking first in the challenge.
Takeaways & Limitations
The extended CDNN outperformed other state-of-the-art methods in the ISBI 2017 skin-lesion segmentation challenge.
Abstract
from arXiv · showhide
Automatic skin lesion segmentation on dermoscopic images is an essential step in computer-aided diagnosis of melanoma. However, this task is challenging due to significant variations of lesion appearances across different patients. This challenge is further exacerbated when dealing with a large amount of image data. In this paper, we extended our previous work by developing a deeper network architecture with smaller kernels to enhance its discriminant capacity. In addition, we explicitly included color information from multiple color spaces to facilitate network training and thus to further improve the segmentation performance. We extensively evaluated our method on the ISBI 2017 skin lesion segmentation challenge. By training with the 2000 challenge training images, our method achieved an average Jaccard Index (JA) of 0.765 on the 600 challenge testing images, which ranked itself in the first place in the challenge
I. INTRODUCTION
The paper addresses difficult, variable dermoscopic lesion segmentation for computer-aided melanoma diagnosis by extending a deep fully convolutional model with architectural and color-input improvements.
- Motivation: Automatic lesion segmentation supports computer-aided melanoma diagnosis but remains difficult across variations in lesion location, shape, size, color, and texture.Low contrast and artifacts such as hairs, frames, blood vessels, and air bubbles further complicate segmentation.
- Motivation: Traditional clustering, thresholding, region-growing, and deformable-model methods have shown limited success on large image datasets.
- Related work: Deep convolutional networks have been adopted for medical image analysis, including lesion detection, registration, segmentation, and classification.
- Contributions: The proposed extension uses a deeper architecture with smaller kernels to increase discriminative capacity across image-acquisition conditions.
- Contributions: It also adds HSV and CIELAB channels alongside RGB and evaluates the framework on the ISBI 2017 challenge.
II. DATASET AND PRE-PROCESSING
The study uses the ISBI 2017 challenge datasets and preprocesses dermoscopic images with resizing and multi-color-space channel construction for CDNN input.
- Dataset: The ISBI 2017 database contains 2000 annotated training images and 600 blind held-out testing images, with an independent 150-image validation set.The training, testing, and validation sets include 374, 117, and 30 melanomas, respectively.
- Color representation: Each color channel can be contrast-normalized to a [5, 95] percentile window, as illustrated by adjusted versions of the original dermoscopic image.
- Pre-processing: Images are resized to 192 × 256 using bilinear interpolation to balance segmentation performance and computational cost.This follows the predominant 3:4 height-to-width ratio in the training set.
- Color representation: The model input combines RGB with three HSV channels and the CIELAB lightness channel, producing 192 × 256 × 7 inputs.These spaces separate image intensity from color information to support more independent channel processing.
III. METHODS
The method is an end-to-end fully convolutional CDNN that processes whole images, reconstructs segmentation maps, and uses a deeper small-kernel design with Jaccard-distance training.
- Network design: CDNN processes an entire image in one pass, using contextual information while avoiding repeated central-pixel classification on small patches.
- Network design: The architecture has convolutional and deconvolutional pathways that extract hierarchical features and reconstruct segmentation maps from coarse to fine resolution.This combines global information with fine details for tumor segmentation.
- Network design: The extended network uses mostly 3 × 3 kernels and additional layers, reaches 29 layers, and contains 5,042,589 trainable parameters.The deeper design preserves the effective receptive field while enabling more nonlinear transformations and reducing the number of weights relative to larger kernels.
- Training objective: The CDNN maps an input image directly to a segmentation map whose elements are pixel-level tumor probabilities.
- Training objective: Training minimizes a Jaccard-distance-based loss using target and output values for each pixel.The loss is directly related to segmentation assessment through the Jaccard index and does not require class re-balancing under foreground-background imbalance.
- Inference: At inference, the resized seven-channel image produces a segmentation map that is converted into a binary tumor mask with dual thresholds.A high threshold of 0.8 identifies the tumor center before a lower threshold is applied.
B. Implementation details
The CDNN was trained from scratch with Adam, dropout, and image augmentation, then evaluated using cross-validation and an ensemble strategy.
- Adam stochastic optimization trained the model from scratch with an initial learning rate of 0.003.
- Two dropout layers with p = 0.5 were added before conv-4-1 and decv-5-1 to reduce overfitting.
- Geometric transformations and per-channel contrast normalization augmented training images to improve robustness across acquisition conditions.
- Five-fold cross validation and grid search were used for model evaluation and hyper-parameter selection.
- A bagging-type ensemble averaged outputs from six models when applying the trained system to testing images.
A. Evaluation metrics
Performance was evaluated by comparing predicted lesion masks with expert-created ground truths using five segmentation metrics, with JA determining the final challenge rank.
- Pixel-wise accuracy, sensitivity, specificity, dice coefficient, and Jaccard index measured segmentation performance against expert ground truths.
- JA = TP/(TP + FN + FP) defines the Jaccard index used for evaluation.
- AC = (TP + TN)/(TP + FP + TN + FN) measures pixel-wise accuracy.
- DI = 2 · TP/(2 · TP + FN + FP) defines the dice coefficient.
- The final challenge ranking was based on JA, with TP, TN, FP, and FN denoting the four pixel-classification counts.
B. Experiments on network architectures
The experiments compared the deeper CDNN-29 with the previous CDNN-19 on validation and testing datasets to assess whether smaller kernels could preserve receptive fields while increasing depth.
- CDNN-29 was compared with the previous CDNN-19 to evaluate the effect of increased depth and smaller convolutional kernels.
- The architecture comparison was first conducted on the validation dataset because testing ground truth was withheld during the challenge.
- The new CDNN achieved better segmentation performance on most metrics across both validation and testing datasets.
- The authors report that increasing depth while reducing filter size can improve CDNN discriminative capability while maintaining an equivalent receptive field.
C. Experiments on input channels
The input-channel experiment compared RGB alone with RGB plus HSV and CIELAB lightness channels, finding narrower training–validation error gaps and improved Jaccard performance with the additional channels.
- Figure 2 compares CDNN training and validation using RGB channels alone versus RGB plus four additional HSV and L channels.
- The additional channels narrowed the gap between training and validation errors during network training.
- Validation error with the additional color channels was consistently lower than with RGB alone.
- 2.4% improvement in Jaccard index occurred on the validation dataset with the additional four input channels.
- 1.5% improvement in Jaccard index occurred on the testing dataset with the additional four input channels.
D. Comparison with other methods in the challenge
On the 600-image ISBI 2017 testing set, the proposed method achieved the challenge’s highest average Jaccard Index.
- 0.765 average Jaccard Index (JA) ranked the method first among challenge submissions.The final challenge included results from 21 participating teams, with the top ten listed in Table VI.
- The evaluation used 600 held-out testing images from the ISBI 2017 skin lesion segmentation challenge.
- Figure 3 presents examples spanning nevus, seborrheic keratosis, and melanoma lesions under varied imaging conditions.
V. DISCUSSIONS
The discussion attributes performance to deeper small-kernel architecture, complementary color-space inputs, and Jaccard-distance training, while acknowledging remaining errors and search limits.
- Remaining weaknesses and future directions: The study could not explore all network configurations because long training times made exhaustive architecture searches infeasible.The ISBI 2017 benchmark provided a common dataset for evaluating and comparing different designs.
- Model design and training: Deeper architecture with small kernels enabled more nonlinear transformations while reducing trainable parameters relative to the shallower model.The authors attribute this design to improved discriminant capability.
- Model design and training: Additional color-space features made training more efficient and robust by explicitly adding complementary information beyond RGB.
- Model design and training: Jaccard-distance loss focused training on segmentation accuracy and handled foreground–background imbalance without class re-balancing.
- Remaining weaknesses and future directions: Some cases still had low segmentation performance, and the authors proposed combining deep models with conventional methods such as active contours.The model’s output correlation was 0.759 with MedMos and 0.873 with NLP LOGIX, suggesting greater complementarity with the active-contour model.
VI. CONCLUSION
The paper concludes that its extended CDNN improves lesion segmentation across varied acquisition conditions by combining deeper small-kernel structures with multiple color spaces.
- The extended CDNN uses deeper network structures with smaller kernels to segment lesions across a wider variety of image acquisition conditions.
- Figure 4 documents cases where the model mis-segmented the wrong sites.
- Figure 5 compares the model with NedMos and NLP LOGIX using scatter plots.
- Combining information from multiple color spaces further boosted segmentation performance.
- The approach outperformed other state-of-the-art methods on the ISBI 2017 skin lesion segmentation challenge.