Machine Learning

We know there is a lot of data generated in the world by computers, people, mobile and other devices. Data is all around us in the form of text, images, videos, music and data is exploding. The volume of data has surpass the human ability to make sense and writing the rules for it manually. 

We are constantly looking at automated systems that can learn the changes in data.

Machine learning is not magic, but tools and technology that can answer our questions with data.

Machine learning grants computers an entire new ability like filtering emails as spam or not spam, face recognition in photos, search recommendations and more. It is changing our day-to-day life.

Machine learning uses data for training to create an accurate model. It can use supervised or unsupervised learning, which depends on labels of data.

Supervised learning is when we have input variables and output variables and we use an algorithm to learn the mapping function between them.

Machine learning algorithms are popular in the medical field in interpreting the medical images. In this article we will talk about a machine learning application: Multiple instance learning.

Multiple Instance Learning

Multiple instance learning is a supervised learning framework in which every training value has a label either discrete or real valued. Multiple instance learning deals with the problem of incomplete knowledge of labels in training sets.

The two primary tasks while applying computer vision in the medical field are image classification for diagnosis and segmentation to separate lesions. In pathology cancer detection, getting labels is often difficult can say time-consuming process. We can use multiple instance learning in such scenarios, it is weakly supervised method which takes a set of labeled bags instead of labeled instances. We do not know the label of each instance and so we need MIL to save labeling efforts and handle weakly labeled data. Here, whole bag needs to be labeled and not each instance. In the below image, the bag is shown containing multiple instances in it.

Each instance xi in bag has a label yi
The label of bag i.e., Y is defined as 

Y=1, if yi==1 (known as positive bag)

Y=0, if for every yi, yi==0 (known as negative bag)

The bag label is assigned value as 0, if all instance labels are 0 and if any instance label is 1 then bag label is 1.

The positive and negative bags can also be shown as below positive instances are shown as green triangle and negative instances as red circle.

Depending on instances in bag, the bag label is assigned. The illustration for the same is shown in the figure below.

MIL uses bags of instances which comprise a label and does the medical diagnosis by predicting the label of the bag. MIL proved to be an ideal solution in medical image analysis and also beneficial in identifying patterns in medical images and scan.

Every image is a bag of instances, and each bag has only one label assigned to it. The goal of MIL is to identify the label of a bag which is a medical diagnosis. I have used the concept of MIL in my final year project for detection of brain tumor detection and the results obtained were outstanding.

Multiple instance learning is new booming method in medical and have shown several best results for histopathological datasets.