Targeted transfer learning to improve performance in small medical physics datasets

TL;DR

This paper reviews and evaluates transfer learning techniques for small medical imaging datasets, highlighting effective strategies like one-cycle training and discriminative learning rates to enhance neural network performance.

Contribution

It identifies optimal transfer learning practices for small datasets and demonstrates the importance of domain-specific transfer learning over generic pre-trained models.

Findings

Transfer learning significantly improves performance with as few as 50 images.

One-cycle training and discriminative learning rates yield best results.

Transfer learning from similar body parts outperforms generic ImageNet pre-training.

Abstract

The growing use of Machine Learning has produced significant advances in many fields. For image-based tasks, however, the use of deep learning remains challenging in small datasets. In this article, we review, evaluate and compare the current state-of-the-art techniques in training neural networks to elucidate which techniques work best for small datasets. We further propose a path forward for the improvement of model accuracy in medical imaging applications. We observed best results from one cycle training, discriminative learning rates with gradual freezing and parameter modification after transfer learning. We also established that when datasets are small, transfer learning plays an important role beyond parameter initialization by reusing previously learned features. Surprisingly we observed that there is little advantage in using pre-trained networks in images from another part of the body compared to Imagenet. On the contrary, if images from the same part of the body are available then transfer learning can produce a significant improvement in performance with as little as 50 images in the training data.