Hand Bone Image Segmentation

Period: 2024.11 · Affiliation: Naver BoostCamp AI Tech 7th (Computer Vision Track)

Background

Bones have a significant impact on the structure and function of our bodies, so accurate bone segmentation is essential for developing medical diagnoses and treatment plans.

Problem Definition

An image containing a hand bone X-ray object is used as the model’s input. The model performs a multi-channel prediction that produces a probability map over 29 classes, and then assigns each pixel to its corresponding class based on this map. Finally, the predicted results are converted into Run-Length Encoding (RLE) format for submission.

Role

Experiments to improve model performance.

Approach

🔗 View individual experiments & contributions

• Attempted returning the loss of an auxiliary classifier
• Searched for the optimal loss suited to the metric and the data
• Attempted pseudo labeling
• Attempted pixel shuffling / unshuffling to compensate for pixels lost due to interpolation
• Attempted adding a 1×1 convolution layer to improve accuracy in the overlapping regions of the back-of-hand bones

Lessons Learned

• The importance of being careful when selecting a model before running experiments
• For some experiments, running follow-up experiments such as weight tuning can be more effective than running a wide variety of experiments
• When working on a project within a limited time, time must be allocated well across the experiments you want to run, and it is worth moving on if an approach shows no effect or remains unsolved within a few days