CombNet: Diagnosis of Skin Diseases using Multi-Scale Features

This is a Pytorch implementation of CombNet. (paper here)
CombNet is a neural network created for image classification: conceived in the feature pyramid structure. I used datasets from kaggle - Skin Cancer MNIST: HAM10000

---

Architecture

Backbone of CombNet is made up of ResNet-18, supported by two types of sub-network from different scale feature maps.
The following formula is used to calculate the total loss from the three losses:

In this experiment, the best result was when alpha = 0.5, betha = 0.7
From this model, you can get around 5% higher accuracy than the plain ResNet accuracy!

Environment

The model is trained using following hardware:

• GTX TITAN X (Pascal) - 12GB VRAM
• Intel(R) Core(TM) i7-5930K CPU @ 3.50GHz
• 64GB RAM

The code is developed under the following software:

• Ubuntu 16.04.6 LTS
• CUDA V10.1.243
• Python 3.6.10
• PyTorch 1.5.0