Skin Lesion Segmentation on ISIC 2018

---
project status: completed baseline + active extensions
---

Overview

DermaSeg is a medical image segmentation project focused on automated skin lesion boundary detection from dermoscopic images. The project is built around the ISIC 2018 Task 1 lesion segmentation benchmark and is designed as a reproducible comparison framework across classical CNN segmentation models, attention-based architectures, lightweight transformer models, and promptable foundation-model baselines.

Rather than training a single model in isolation, DermaSeg is structured as a complete segmentation workflow: dataset loading, preprocessing, supervised training, validation-based checkpoint selection, test-set evaluation, qualitative visualization, and metric summarization.

Current Status: Strong ISIC Baseline

The current completed experiment uses a DeepLabV3 model with a pretrained ResNet-50 backbone for binary lesion-vs-background segmentation.

These results come from a local run on the official ISIC 2018 split, not from copied leaderboard values. The best checkpoint is selected using validation threshold Jaccard and then evaluated on the held-out test set.

Technical Approach

The project uses a supervised 2-class segmentation setup:

• Dataset: ISIC 2018 Task 1 lesion boundary segmentation.
• Input: RGB dermoscopic images resized to 320 × 320.
• Model: DeepLabV3 with a pretrained ResNet-50 backbone.
• Loss: Combined cross-entropy and Dice loss.
• Selection Metric: Validation threshold Jaccard.
• Outputs: Saved checkpoints, metrics JSON files, Markdown experiment summaries, and qualitative prediction panels.

DeepLabV3 was chosen as the first strong baseline because it provides a serious non-U-Net comparison point. Its atrous convolutions and ASPP module allow the model to capture multi-scale lesion context, which is useful for irregular lesion boundaries and variable lesion sizes.

Model Coverage

DermaSeg is designed to compare multiple segmentation model families under a shared training and evaluation pipeline.

Supervised Segmentation Models

• U-Net
• Attention U-Net
• SegNet
• U-Net++
• DeepLabV3
• SwinUNetLite
• SegFormerLite

Promptable Foundation-Model Extensions

• SAM
• MedSAM
• Box-prompted evaluation workflow
• Separate promptable evaluation path to avoid mixing supervised and foundation-model comparisons unfairly

Qualitative Evaluation

The repository also exports qualitative prediction panels from the saved DeepLabV3 checkpoint. Each panel compares:

1. the original dermoscopic input image,
2. the ground-truth lesion mask overlay, and
3. the predicted lesion mask overlay.

This makes the project more interpretable than a metrics-only benchmark and allows direct inspection of boundary quality, under-segmentation, over-segmentation, and lesion-shape errors.

Why This Project Matters

DermaSeg demonstrates practical medical image segmentation beyond a one-off notebook. It shows that the same dataset and evaluation protocol can be used to compare classical encoder-decoder CNNs, atrous-convolution models, transformer-style architectures, and promptable segmentation models.

For my broader work in medical imaging and computer vision, this project strengthens my experience with:

• reproducible segmentation pipelines,
• medical benchmark datasets,
• lesion boundary detection,
• Dice/IoU/Jaccard-based evaluation,
• qualitative model inspection,
• CNN and transformer segmentation architectures,
• and foundation-model extensions for medical image analysis.

Next Steps

The next phase of DermaSeg will focus on making the benchmark more complete and clinically meaningful:

1. Loss Ablations: Compare cross-entropy + Dice, Tversky, and focal Tversky losses for imbalanced lesion segmentation.
2. Architecture Benchmarking: Run full evaluations across U-Net, Attention U-Net, U-Net++, DeepLabV3, SwinUNetLite, and SegFormerLite.
3. Promptable Segmentation: Evaluate SAM and MedSAM using box prompts derived from lesion masks.
4. Failure-Mode Analysis: Categorize segmentation errors by lesion size, contrast, boundary ambiguity, and image artifacts.
5. Portfolio Visualization: Add qualitative result panels directly to the project page for visual comparison across models.