[Paper Review] II-FCN for skin lesion analysis towards melanoma detection

Hongdiao Wen|arXiv (Cornell University)|Feb 28, 2017

Cutaneous Melanoma Detection and Management8 references17 citations

TL;DR

This paper proposes II-FCN, a symmetric fully convolutional network with identity inception blocks for melanoma segmentation in dermoscopy images. It integrates dilation convolution, conditional random fields (CRF), and a novel approximate Jaccard index loss with loss normalization to improve boundary accuracy and generalization, achieving a Jaccard index of 0.82 on the ISBI 2017 dataset without overfitting.

ABSTRACT

Dermoscopy image detection stays a tough task due to the weak distinguishable property of the object.Although the deep convolution neural network signifigantly boosted the performance on prevelance computer vision tasks in recent years,there remains a room to explore more robust and precise models to the problem of low contrast image segmentation.Towards the challenge of Lesion Segmentation in ISBI 2017,we built a symmetrical identity inception fully convolution network which is based on only 10 reversible inception blocks,every block composed of four convolution branches with combination of different layer depth and kernel size to extract sundry semantic features.Then we proposed an approximate loss function for jaccard index metrics to train our model.To overcome the drawbacks of traditional convolution,we adopted the dilation convolution and conditional random field method to rectify our segmentation.We also introduced multiple ways to prevent the problem of overfitting.The experimental results shows that our model achived jaccard index of 0.82 and kept learning from epoch to epoch.

Motivation & Objective

• To address the challenge of low-contrast, poorly distinguishable skin lesions in dermoscopy images for melanoma detection.
• To develop a lightweight, robust fully convolutional network that generalizes well despite limited training data.
• To improve segmentation accuracy by integrating advanced components such as dilation convolution and CRF for boundary refinement.
• To overcome overfitting in deep networks through dynamic data augmentation and loss normalization.
• To propose a differentiable, approximate Jaccard index loss for end-to-end training that aligns with the evaluation metric.

Proposed method

• Proposes a symmetric identity inception block composed of four convolution branches with varying kernel sizes and depths, enabling feature fusion without zero padding.
• Employs a reversible architecture using transposed convolutions to allow symmetric encoder-decoder design with reduced parameters.
• Introduces a normalized loss function that reweights foreground and background pixels based on object percentage to balance class imbalance.
• Designs an approximate differentiable Jaccard index loss using a logarithmic approximation to optimize for the evaluation metric directly.
• Applies dilation convolution in the final layers to expand receptive field and eliminate 'black holes' in segmentation.
• Uses conditional random fields (CRF) as a post-processing step to refine boundaries and suppress small spurious segments.

Experimental results

Research questions

• RQ1 Can a symmetric, lightweight fully convolutional network with identity inception blocks outperform standard FCN architectures in skin lesion segmentation?
• RQ2 Does combining dynamic data augmentation with loss normalization improve model generalization and reduce overfitting?
• RQ3 To what extent does an approximate Jaccard index loss improve training alignment with the evaluation metric?
• RQ4 How effective is the integration of dilation convolution and CRF in enhancing boundary precision and reducing false positives?
• RQ5 Can a model trained from scratch achieve high performance without transfer learning, using only data augmentation and loss engineering?

Key findings

• The II-FCN model achieved a Jaccard index of 0.82 on the ISBI 2017 dermoscopy lesion segmentation benchmark.
• The model maintained consistent learning progress across epochs, indicating effective training and generalization.
• The CRF post-processing step improved the baseline Jaccard index by an average of 0.005.
• The combination of dynamic data augmentation and loss normalization reduced the discrepancy between training and test data distributions.
• The model demonstrated robustness to overfitting despite training from scratch with limited data, thanks to architectural and training innovations.
• The approximate Jaccard index loss enabled effective optimization toward the primary evaluation metric, showing strong alignment with final performance.