Conclusive statements by Allwyn Team on restructuring the new ML algorithm for L3 detection

So far, in this series of posts, we walked you through the problem statement, objectives, methodology, and approach of how we performed the restructuring/redefining of the existing L3 detection algorithm with the help of Machine Learning.

SUMMARY OF RESULTS

A validation set was created to determine the effects of modifying the pre-processing parameter. It contained images from 5 collections. Preliminary results are summarized in a confusion matrix to represent the recommended manual review rate (“yes” vs. “review”). Our findings indicate that the algorithm improves the accuracy and efficiency of locating L3 with a high level of confidence.

LESSONS LEARNED

After the team ran the data preparation code, orientation and duplication issues were encountered when running the detection algorithm was attempted. Much time and effort went into ensuring that various types of CT scans could be prepared for analysis; however, without the proper preparation, the analysis is not meaningful.

Separating the data prep and detection helped with the overall workflow and enabled the team to properly ready the data for analysis.

The team had a slow start in implementing improvements because there was a long lead time to understand the code.

We have commented on both the code and the GoLab notebook. We have also created a procedure in our report so that other users can quickly begin implementation.

FUTURE WORK

We found examples in our dataset of CT scans that are not performed top-down, but rather side-to-side. Future studies should focus on ensuring that the MIP can create adequate 2D images from these types of scans.

Since every image is different, a one-size-fits-all approach to pre-processing may not work. It may be necessary to vary the HU scale – as well as other pre-processing parameters – to choose optimal parameters and analyze the

effect on confidence.

We plan to use the existing Convolutional Neural Network (CNN) and train it with new labeled data to identify other features of interest. We will also study the use of CT scans performed side-to-side.

Conclusion

A machine learning algorithm was created that successfully identified LC slices from full-body CT scans. The team has used an existing algorithm to detect L3 slices in CT scans and improved it by:

  • Separating the pre-processing and detection code for a more manageable workflow
  • Organizing both pre-processing and detection outputs into an easily understandable and navigable file structure
  • Adjusting pre-processing to improve the detection rate

We have brought value to the problem space by:

  • Creating a codebase that is easy to navigate and allows for multiple detection runs
  • Using pre-processing as the main method of improving detection
  • Identifying the algorithm’s weak points with a validation set that had previously unknown scenarios
  • Demonstrating a high level of confidence in the AI-based calculations

We hope that our project experience will directly or indirectly help Federal/ State Healthcare agencies and other medical and research organizations. Follow us on LinkedIn to Stay tuned with our upcoming projects helping the US Healthcare industry.