Mun Fai 文辉

Summary

• Published paper in MDPI Journal of Imaging as part of Masters group project
• Won best poster award (500 pounds) out of 40 PHD teams in Imperial College London research showcase

Abstract

Self-supervised learning approaches have seen success transferring between similar medical imaging datasets, however there has been no large scale attempt to compare the transferability of self-supervised models against each other on medical images. In this study, we compare the generalisability of seven self-supervised models, two of which were trained in-domain, against supervised baselines across eight different medical datasets. We find that ImageNet pretrained self-supervised models are more generalisable than their supervised counterparts, scoring up to 10% better on medical classification tasks. The two in-domain pretrained models outperformed other models by over 20% on in-domain tasks, however they suffered significant loss of accuracy on all other tasks. Our investigation of the feature representations suggests that this trend may be due to the models learning to focus too heavily on specific areas.

Please explain to me what all of this means!

Most advances in Machine Learning has been in the field of supervised learning, where data points come with their corresponding labels which the model can use to evaluate its own performance. E.g. a chest X-ray dataset will contain not only images, but labels indicating the types of diseases depicted.

However, the cost of annotating labels is expensive, especially in certain niche contexts like medical imaging. On the other hand, there is an abundance of unlabelled data which we should leverage on. One such solution is self-supervised learning (SSL). Typically, this involves 2 stages.

1. Pretraining:   Here, the model can only access unlabelled data. It learns structural patterns within the dataset.
   • E.g. one popular SSL technique is to give a model an image and an augmented version of the first original image. It is also fed other images, with the goal of learning that the original and augmented image are the same, and different from the other images.
2. Finetuning:   The model is then finetuned on the specific domain we are interested in (and that does not have to be in exactly the same domain as the pretraining phase). If this is a supervised learning problem, our finetuning stage needs to be supervised as well.

In practice, it is too expensive (as it takes weeks) to pretrain one own SSL model from scratch. Very often, one finds a readily available pretrained model and then finetunes it on his/her own purposes, even when the pretraining and finetuning domain may not be exactly the same. In light of this, this paper investigates how well pretrained SSL methods can be used in medical imaging.