Source-linked AI summary
ISIC 2017 - Skin Lesion Analysis Towards Melanoma Detection
Matt Berseth
TL;DR
The paper addresses lesion segmentation and classification in the ISIC 2017 challenge using deep convolutional networks. It uses U-Net for pixelwise segmentation and AlexNet for image classification, while reporting fold-wise Jaccard Index results and identifying potential data leaks in the images.
Problem
The paper addresses the ISIC 2017 challenge tasks of lesion segmentation and lesion classification.
Method
The system uses a modified U-Net to produce same-sized pixel probability maps for segmentation and an AlexNet-based network to classify lesion images.
Results
The evaluation reports the best Jaccard Index score for each validation fold for both lesion segmentation and lesion classification.
Takeaways & Limitations
The workflow combines pixel-level lesion probability estimation with image-level label prediction and removes identified background artifacts from training images.
Takeaways & Limitations
Potential data leaks include colored gauze and bright edge lighting associated with particular classes, with a possible bright light leak also suspected in test images.
Abstract
from arXiv · showhide
Our system addresses Part 1, Lesion Segmentation and Part 3, Lesion Classification of the ISIC 2017 challenge. Both algorithms make use of deep convolutional networks to achieve the challenge objective.
LESION SEGMENTATION
The segmentation system expands and transforms lesion images and masks, then trains a U-Net to produce pixelwise probability maps. Ten-fold training and probability-map averaging support the final segmentation, with models selected by validation Jaccard Index.
- Preprocessing: Training images were resized to 192 × 192 pixels and augmented with elastic distortions, rotations, flips, and zooms.Each original image generated additional distorted examples, while masks underwent matching transformations.
- Model Architecture: The U-Net receives a 192 × 192 image and produces a probability map with the same dimensions.The architecture uses three down-sampling and three up-sampling layers, with a fully connected layer at the bottom of the U.
- Training: Training used Adam with a fixed 1e-4 learning rate, minibatches of 20, and a custom weight map balancing positive and negative classes.Ten-fold cross-validation assigned each original image and its distortions to one fold to prevent leakage.
- Training: The best model for each fold was selected by validation Jaccard Index after 200 epochs.The recorded validation score determined which iteration was retained for each fold.
- Postprocessing: Final probability maps were averaged across the ten fold models before scoring validation and test sets.A conditional random field was tested for refinement but discarded because it did not add significant performance.
- Results: The reported segmentation results present the best Jaccard Index score for each validation fold.The results are summarized in the cross-validation table.
LESION CLASSIFICATION
The classification system augments and crops lesion images, removes identified background artifacts, and uses AlexNet with demographic information in a random-forest postprocessor. Its evaluation uses stratified ten-fold training and validation AUC model selection, while the authors flag possible data leaks in the dataset.
- Preprocessing: Images were cropped to square dimensions, resized to 256 × 256 pixels, and augmented with 90- and 270-degree rotations, producing 6,000 training images.The preprocessing expands the original training set through rotation before resizing.
- Data Quality: Two potential data leaks were identified: visible colored gauze and bright light on both image edges were each associated with a single class.The authors illustrate these patterns in the training-image figures.
- Preprocessing: The affected training images were manually cropped to remove background objects so the model would learn lesion rather than artifact features.This preprocessing specifically targeted the identified gauze and bright-light artifacts.
- Model Architecture: AlexNet classified 224 × 224 crops into label probabilities, using 1024-neuron fully connected layers instead of the traditional 4096.Dropout and rectified linear units were also used in the modified architecture.
- Training: Training used Adam with a fixed 1e-5 learning rate, balanced minibatches of 64, and stratified ten-fold cross-validation.Original and rotated images remained within the same fold to prevent train-test leakage.
- Training and Postprocessing: Models were selected by validation AUC over 300 epochs, and fold probabilities were averaged before combining them with patient demographics in a random forest.The random forest generated the final image probabilities.
- Results: The reported classification results are based only on visual image characteristics and are summarized by cross-validation fold.The section references Table 2 for the fold-level results.
- Limitations: The authors note that bright light leakage may exist in the test set and could affect the final competition metric.They propose recomputing the metric without potentially affected examples if the suspected leak is real.