Hi, I am a graduate student and was looking for ideas for computer vision-based projects. I have rounded up on projects such as
COVID-19 Biomarker Detection in Chest X-rays. But I'd like to know what sort of novelty I can bring into this work? Or can I use another ML algorithm to do the same? Also, it would be really helpful if somebody can tell other ideas. Thank you.

 

using thermal imaging is fast but rather poor method to identify virus or viral load. also it is something with no intrinsic checks to ensure accuracy. for example person that was just infected may have no symptoms but virus is rapidly multiplying and still able to infect others. also a healthy person that was just racing to be someplace on time may have elevated temperature and thermal image profile matching one of few predefined categories even though that is not the case.
further, people using such system to spot and isolate someone that was identified as Covid carrier would have little recourse to verify the results. this is rally a rather poor (band-aid) solution meant only to aid with speed where volume of test subjects is present and medical diagnosis is too slow. false diagnostics at the airport would go undetected for long enough to make any efforts to deal with pandemic completely pointless.

if you are interested in using vision, consider applications that are in demand such as machine guidance (robots, vehicles, bin picking etc.). if your system produces poor result here, it will be very obvious and one will quickly get a chance to work on a system improvement. but once robust, this would have wide application potential

 

Are you allowed to use the Raspberry Pi? If yes, you can use openCV library and try something like this: https://core-electronics.com.au/guides/object-identify-raspberry-pi/

 

So, just because it gives me a chance to cite colleagues at the university from which I retired, here’s a possible direction:

D. H. Ye, S. Srivastava, J. -B. Thibault, K. Sauer and C. Bouman, "Deep Residual Learning for Model-Based Iterative CT Reconstruction Using Plug-and-Play Framework," 2018 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 2018, pp. 6668-6672, doi: 10.1109/ICASSP.2018.8461408.

Abstract: Model-Based Iterative Reconstruction (MBIR) has shown promising results in clinical studies as they allow significant dose reduction during CT scans while maintaining the diagnostic image quality. MBIR improves the image quality over analytical reconstruction by modeling both the sensor (e.g., forward model) and the image being reconstructed (e.g., prior model). While the forward model is typically based on the physics of the sensor, accurate prior modeling remains a challenging problem. Markov Random Field (MRF) has been widely used as prior models in MBIR due to simple structure, but they cannot completely capture the subtle characteristics of complex images. To tackle this challenge, we generate a prior model by learning the desirable image property from a large dataset. Toward this, we use Plug-and-Play (PnP) framework which decouples the forward model and the prior model in the optimization procedure, replacing the prior model optimization by a image denoising operator. Then, we adopt the state-of-the-art deep residual learning for the image denoising operator which represents the prior model in MBIR. Experimental results on real CT scans demonstrate that our PnP MBIR with deep residual learning prior significantly reduces the noise and artifacts compared to analytical reconstruction and standard MBIR with MRF prior.

The research is using ML for denoising in a sort of pipeline configuration. Maybe some ideas there…

 

So, just because it gives me a chance to cite colleagues at the university from which I retired, here’s a possible direction:

D. H. Ye, S. Srivastava, J. -B. Thibault, K. Sauer and C. Bouman, "Deep Residual Learning for Model-Based Iterative CT Reconstruction Using Plug-and-Play Framework," 2018 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 2018, pp. 6668-6672, doi: 10.1109/ICASSP.2018.8461408.

Abstract: Model-Based Iterative Reconstruction (MBIR) has shown promising results in clinical studies as they allow significant dose reduction during CT scans while maintaining the diagnostic image quality. MBIR improves the image quality over analytical reconstruction by modeling both the sensor (e.g., forward model) and the image being reconstructed (e.g., prior model). While the forward model is typically based on the physics of the sensor, accurate prior modeling remains a challenging problem. Markov Random Field (MRF) has been widely used as prior models in MBIR due to simple structure, but they cannot completely capture the subtle characteristics of complex images. To tackle this challenge, we generate a prior model by learning the desirable image property from a large dataset. Toward this, we use Plug-and-Play (PnP) framework which decouples the forward model and the prior model in the optimization procedure, replacing the prior model optimization by a image denoising operator. Then, we adopt the state-of-the-art deep residual learning for the image denoising operator which represents the prior model in MBIR. Experimental results on real CT scans demonstrate that our PnP MBIR with deep residual learning prior significantly reduces the noise and artifacts compared to analytical reconstruction and standard MBIR with MRF prior.

The research is using ML for denoising in a sort of pipeline configuration. Maybe some ideas there…

I am hardly qualified to comment but it seems an overreliance on denoising i.e. is noise the only factor considered?

 

I am hardly qualified to comment but it seems an overreliance on denoising i.e. is noise the only factor considered?

It relies on Machine Learning and the construction of an extensive model. The “replacement” is a ML-based system similar to the very successful Stable Diffusion model which uses denoising as the method to derive the image associated with the prompt.

The PnP idea is that the ML component is equivalent to a denoising operation from the point of view of the combined forward and prior model system in a way that lets the forward model be created without reliance on the content of the prior model.