Project: Low Powered Models for disease detection and classification for Radiology Images


Current mobile/web applications related to detection and classification of images are mostly based on client-server architecture, where the image/data is passed to server which performs the inference on data and returns the results back to application. Another scenario can be that current deep models require high end gpus to train and infer. We would like to have cost-effective and efficient model which can precisely detect and classify diseases from images provided and should be capable of running on mobile devices. It would be helpful to radiologist as well as patient which can see results on the go using their mobile itself.

Students are required to create (or research existing) deep models to detect and classify radiology images using low powered devices like mobile.

Required Skills : ML, DL, Python, Qemu (Hardware emulator to emulate low powered boards)

Mentors : @pri2si17 @sunbiz @judywawira

Deliverables : Students should make a POC to run and deep model on mobile (possibly emulator) and explore other optimization techniques for deep models.

Note : More details will follow in coming days, but students can explore it and come up with their opinions/solutions.


Hi Priyanshu, I am Kunal Mehta, currently in my final year of Computer Engineering at K.J Somaiya. I have worked on many ML and DL projects and would like to work on this topic and contribute to the open-source as much as I can.

1 Like

Hi @kunakl07, welcome to the community. You can make a POC. I will update the instructions shortly. Till then do research of existing models and datasets.

Hi Priyanshu, I am Shreya Goyal, currently pursuing master’s in Health Informatics from IUPUI. I have a strong background in biomedical data analysis and Machine Learning. I believe I have the required skills to contribute to the project and to the organization.

1 Like

Hi Priyanshu (@pri2si17),
I’m Shushant Kumar, BTech final year undergraduate in Computer Science and Engineering from National Institute of Technology Karnataka, India. My prime area of interest has been towards Deep Learning and medical image analysis to transform the data into clinical insights. My major research has been in the field of developing novel models for classification and segmentation of lesions and other specific artifacts in radiology images. Also with my adept experience of open source application development and previous work on designing of lighter models for resource constrained devices, I would like to use this platform to contribute to the project and be a part of LibreHealth community.

Hi @Shreya, welcome to the community. Well surely you can contribute to this project. Please make a POC and let us know if you have any queries.

Hi @shushantkmr2 welcome to the community, please develop a PO and contribute. :slight_smile:

For this if we need the model to be processed on mobile instead of processing on servers we can use flite tensorflow lite for making model in mobile.:grinning: please submit your views. @pri2si17

Hi Priyanshu @pri2si17 I am Arihant Jain , 3rd year CSE undergraduate at Manipal Institute of Technology , India. I have used deep learning to build models for image segmentation of Brain Tumours (MRI) and also classification of Subarachnoid hemorrhage on CT images. Growing up as the son of two doctors has helped me work closely with medical problems. I would like to learn and contribute to LibreHealth community . This will indeed allow me to help a lot of people. Looking forward to working with everyone. :smiley:

Hello @pri2si17. I am Rupal Sharma, an undergraduate interested in this project. I have a query regarding POC and project. Is it necessary to make the model for mobile devices i.e., can we make app which will run on browser using tenosorflowjs? Thanks in advance.

Hi @rsdel2007, welcome to the community. You can make a web app, it is not at all desired to have mobile application for POC however making such is a plus. The constraint you will face is that you will be using your laptop resources which are pretty high as compared to your mobile device. I would suggest you to have a simple mobile application with web view, where you can actually see your model running on resource constraint device. A simple emulator will suffice the need. Let us know if you have any issues.

Hi @ajarihantjain54, welcome to the community. Please develop a POC. :slight_smile:

Hi @gauthampkrishnan, there is a catch. Tensorflow lite is no doubt for mobile devices, but your accuracy/performance will be impacted a lot. Why not develop on PC and optimize it to work on mobile? I guess I have given you the hint. :slight_smile:

Yes that’s possible, we can do like train using the computer and we can push a update every 2 months. This makes sure that the app is updated with an optimised model. And this can makes sure that we dont need to use any mobile processing power too. :slightly_smiling_face:

I am Ritwick Ghosh. I am interested in GSoC2020 as student. According to my experience, for mobile application type of use we have to optimize 3 prospectives for a CNN model:

  1. Size of model.
  2. Parameters of the model.
  3. Evaluation time of model. These three are not exclusive, but may differ according to the preference of model network and architecture. Please add additional details. I would love to work on this.

Hi @RitwickGhosh, You are correct. Please make one such small demo. I will update the necessary details accordingly. You have lot of stuffs to explore for mobile network :slight_smile:

@pri2si17 Ok, I am on it. Update you soon.

@pri2si17 what if we create a model and pass it into the inference engine to get optimized Edge App

Hi @shashank, this will defeat the purpose. We don’t want client-server setup, model should run on mobile platform itself.

1 Like

Remember: infrastructure in a lot of areas is non-existent, very distant, or just plain horrible.