Skin Lesion Segmentation Model Development
Budget / Salary₹1,500–12,500
TypeFreelance project
LocationRemote
Posted1 hour ago
I am building a deep-learning pipeline that can draw an accurate outline around suspicious skin lesions. Right now I only have images tagged as cancerous or benign—no pixel-level masks—so I need help turning those image-level labels into a reliable segmentation model.
Here is what I’m aiming for:
• A model that returns a precise outline following the lesion’s actual shape, not just a bounding box or heatmap.
• Training that makes the most of cloud GPUs or other efficient strategies, since I don’t have local high-performance hardware.
• A workflow I can reproduce later—ideally scripted in PyTorch or TensorFlow, containerised with Docker, and runnable on Colab, AWS, or a similar platform.
I’m open to weakly supervised or semi-supervised techniques that can leverage my label-only dataset, as well as any annotation tricks you suggest to bootstrap accurate masks quickly. Explainable AI add-ons (e.g., Grad-CAM visualisations) would be a welcome bonus, but the primary deliverable is a well-trained segmentation network—U-Net, DeepLab, or whatever architecture you feel suits the task—that outputs tight lesion contours.
Deliverables
• Code repository with training, validation, and inference scripts
• Trained model weights and sample inference results
• Clear setup and run instructions, including cloud GPU configuration
• Brief report on the method, data preprocessing, and metrics used (IoU/Dice preferred)
If you’ve built medical-image segmentation models before and can guide me through cloud deployment while keeping compute costs reasonable, I’d love to collaborate.
Here is what I’m aiming for:
• A model that returns a precise outline following the lesion’s actual shape, not just a bounding box or heatmap.
• Training that makes the most of cloud GPUs or other efficient strategies, since I don’t have local high-performance hardware.
• A workflow I can reproduce later—ideally scripted in PyTorch or TensorFlow, containerised with Docker, and runnable on Colab, AWS, or a similar platform.
I’m open to weakly supervised or semi-supervised techniques that can leverage my label-only dataset, as well as any annotation tricks you suggest to bootstrap accurate masks quickly. Explainable AI add-ons (e.g., Grad-CAM visualisations) would be a welcome bonus, but the primary deliverable is a well-trained segmentation network—U-Net, DeepLab, or whatever architecture you feel suits the task—that outputs tight lesion contours.
Deliverables
• Code repository with training, validation, and inference scripts
• Trained model weights and sample inference results
• Clear setup and run instructions, including cloud GPU configuration
• Brief report on the method, data preprocessing, and metrics used (IoU/Dice preferred)
If you’ve built medical-image segmentation models before and can guide me through cloud deployment while keeping compute costs reasonable, I’d love to collaborate.
Apply on Freelancer →
Project sourced from Freelancer.com. Applications happen directly on the original platform — we never collect your data.