Abstract
Challenge: Fractures are local patterns that require local analysis. But hospitals often don't specify areas of interest.
a two-stage hip and pelvic fracture detection method that executes localized fracture classification using weakly supervised ROI mining
A two-stage hip and pelvic fracture detection method using weakly supervised ROI mining for local fracture classification
1st: Large-capacity multi-instance fully convolutional network ⇒ Mining possible real samples and determining difficult negative ROIs
2nd: Small-capacity model ⇒ Locally analyze the mined ROIs and classify fractures (the second stage screens the false positives in the first stage)
During inference: Connect the output of the two stages to detect hip-pelvic fractures at once
4410 pictures, hip and pelvis tested together
ImageNet pre-trained two-stage model
More effective than doctors (although not as good as experts), they are the first to diagnose patients
Previous work only focused on the diagnosis of hip fractures and did not classify more complex pelvic fractures (3 pelvic fractures: ilium, ischium, and pubis).
We are the first to address automated PXR pelvic fracture classification and the first to demonstrate the diagnostic capabilities of hip and human physicians for pelvic fractures.
[Supplementary knowledge points]
Weakly Supervised Learning
Weakly Supervised Learning ☆: In the field of machine learning, learning tasks can be roughly divided into two categories, one is supervised learning and the other is unsupervised learning. Typically, both require learning a predictive model from a training dataset containing a large number of training samples, each corresponding to an event/object.
Although current supervised learning technology has achieved great success, it is worth noting that due to the high cost of the data annotation process, it is difficult to obtain strong supervision information such as all true value labels for many tasks. However, unsupervised learning has developed slowly because the learning process is too difficult. Therefore, it is expected that machine learning techniques can work under weak supervision. Professor Zhou Zhihua of Nanjing University published a paper called "A Brief Introduction to Weakly Supervised Learning" in January 2018, which gave a new trend and ideas for machine learning tasks.
Weakly supervised learning can be divided into three typical types:
- Incomplete supervision means that only part of the training data is given labels , and some data is not labeled.
- Inexact supervision means that the training data only provides coarse-grained labels . We can think of the input as a bag. There are some examples in this bag. We only know the label of this bag, Y or N, but we don't know the label of each example.
- Inaccurate supervision means that the labels given are not always correct . For example, the label that should be Y is incorrectly labeled as N.
- We will treat these types of weakly supervised learning separately, but it is worth mentioning that in practice they often occur simultaneously. [The specific solution ideas are illustrated in the article with examples]
In order to solve incomplete supervision, we can consider two main techniques, Active learning and Semi-supervised learning. One is with human intervention, and the other is without human intervention.
In order to solve the problem of inaccurate supervision, we can consider multi-instance learning.
In order to solve the problem of inaccurate supervision, we consider Learning with label noise.
In the 2019 CVPR, there was also a lot of progress in weakly supervised learning. Weak supervision is summarized as follows:
