Coach Meeting – 12/4/23
Key Takeaways
- Model has finished training and image visualization samples have been shared with client
- Something to think about is implementing allowing a direct download of source code rather than requiring cloning the repository
Next Steps
- Integrate a simple command line interface for client to use model
- Pull out handoff plan into its own section/page and expand on what we currently have
- Give specifics in handoff plan about next steps for model and how to continue iteration
Team Meeting – 11/26/23
Key Takeaways
- Model had training time error (too many epochs, so have to decrease the number)
- Training model again with less epochs, should finish training on 11/29/23
Next Steps
- Wait for model to finish training, create visualization code for the output images
- Finish the handoff plan
Coach Meeting – 11/17/23
Key Takeaways
- Could pass some files into the command line – maybe contained within a README
- Two different models are currently being trained, will take 3-6 days for these models to be completely trained
Next Steps
- Find time with client to meet and demo what we have
- Finish the documentation plan
Coach Meeting – 11/10/23
Key Takeaways
- APPLES reflection has been completed
- Demo shown – model outputs with specific labels for model with 30 training epochs
Next Steps
- Continue training model and creating other models
- Show client the demo we have prepared
Team Meeting – 11/06/23
Key Takeaways
- Team set-up LongLeaf for running the model
- Team uploaded the data into LongLeaf to be used for model training
Next Steps
- Run the model in LongLeaf for training and testing
Team Meeting – 11/04/23
Key Takeaways
- Team set-up Colab notebook and began coding model
- Data loading and pre-processing
- Downloaded necessary datasets (BraTS 2020) – Brain Tumor MRIs
- Scheduled meeting with client for later this week (Thursday most likely)
Next Steps
- Meet tomorrow to continue model development
- Finalize meeting time with client this week
Coach Meeting – 11/03/23
Key Takeaways
- Team updated Ethics assignment with specific examples for each section
- Team is currently working on model development with hopes of demo in the coming 2 weeks
- Preparation for Tech Talk on Tuesday, slides are completed and demo is prepared
Next Steps
- Meet with client after initial model development for touchpoint and adjustments
- Continue working on model development
Coach Meeting – 10/27/23
Key Takeaways
- Ethics assignment should include specific examples (i.e. companies/softwares)
- Update the architecture diagram as we learn more about specifications
- Earlier we can demo, the easier it will be for us
Next Steps
- Within 2 weeks from today, having a fully functioning model that can segment ready for demo
- Meet with client to discover more about CLI vs UI requirements
Team Meeting – 10/17/23
Key Takeaways
- Met to set up Docker images to transfer into the Longleaf cluster
- Converted Docker image into Singularity image – to be used by Longleaf
- Worked on backend code development in Collab to be trained in Longleaf
Next Steps
- Continue working on Longleaf cluster to run previous code
- Continue working on backend code development
Client Meeting 3 – 10/11/23
Key Takeaways
- Team is caught up with GAN and model content
- Koushik has LongLeaf access now
- The pip file that Andrew shared with us should contain all of the dependencies
- Front-end should
- Allow upload of folder, zip, or images
- Will process through the GAN model
- Return the segmented images
Next Steps
- Have the Docker instance set-up and run the initial code Andrew sent to better understand
- Install Python in Docker
- Pip Install
- Switch current code to function with CUDA
- Understand current code and functions
Client Meeting 2 – 09/25/23
Key Takeaways
- Andrew uploaded initial code for piping in CT scans and semantic segmentation
- Will need to set up a python virtual environment
- Sent an email to our professor about access to LongLeaf and other computing resources from the University.
- Could also concurrently setup LongLeaf since it is simple and easy.
- Reasoning for not using pre-trained GAN models?
- There exists pre-trained segmentation model, mainly for the abdomen, but when Andrew tried to train these models on the brain, but there’s no control over the number of layers, parameters, etc…
- Chose to set-up a tensorflow because they were slow and did not support multi-GPU processing.
Next Steps
- Send Andrew a single email with everyone’s dropbox email for the CT scan data
- BRATS data for brain MRIs
- Will look at the code that Andrew has sent over so that we can understand implementation for CT scans.
- Setup the anaconda virtual environment to run the code and have the dependencies
Coach Meeting 2 – 09/22/23
Key Takeaways
- User Stories should be more broad and be less revealing of the implementation details
- More user interactions with less focus on technical aspects
- Discussed getting access to LongLeaf, should follow up with Stotts for more information
Next Steps
- Remove technical details from User Stories
- Meet with client to discuss action items and deliverables to complete
Team Meeting 2 – 09/20/23
Key Takeaways
- Team chose to proceed with Brain MRI for our model
- Team Roles
- Luke will handle front-end and help out with back-end if needed
- Koushik and Daniel will handle main model development
- Krishna to assist with both front-end and back-end development
- Team completed D1 Specifications and began to look at Ethics assignment
Next Steps
- Questions for Andrew:
- Since there are already medical GAN models that segment images, what is stopping us from using these models?
- Where do we want to host the web-app?
- Tarheel
- Firebase
- GitHub pages
Coach Meeting 1 – 09/15/23
Key Takeaways
- Deliverables to be completed the Sunday following the classes/week written down for
- First client meeting set for today at 4:30PM
- Should relay team understanding of project/scope to ensure alignment with client requirements
- Can detail implementation point priorities in user stories
Next Steps
- After client meeting, decide deliverables and roles going forth in the project
- Finish user stories and project concept to be completed and posted on website
Client Meeting 1 – 09/15/23
Key Takeaways
- All current code is written in Python with API calls
- Setting up an anaconda environment on Macs (or windows) should be good
- Focused on semantic segmentation for CT scans – skin, bone, tissue, etc…
- Previously had to manually segment CT scans by hand (took too long and tedious task)
- Not most efficient training, low sample size
- Could train GAN for very specific objects in a CT scan (intracranial hemorrhages, tumors, etc…)
- Essentially our task is to create a “neurosurgeon” to label the CT scans
- Initially the model will take a CT scan and be able to label the sections (blood, skin, brain, bone, etc…)
- Head CT is 512 images each 1024 x 1024
- Main project objectives:
- Set up the GAN
- Input Brain CT to the model for segmentation
- For running/training the model, can either run on:
- University supercomputer
- Andrew has an extra 3090
- Could buy a shared network
- Google Collab
Next Steps
- Send email to professor to inquire about possible usage of LongLeaf or Collab
- Decide if we want to do Brain (CT or MRI) or Spine
- Do more readings/research on GANs and usage (TensorFlow Documentation)
- Moving forward, weekly meetings and more/less as necessary
- Primary mode of conversation to be text
09/08/2023 – Team Meeting 1
Key Takeaways
- Will make specific technical (tools, packages, etc…) and structural (project roles) decisions after initial client meeting
- Daniel Fang will take lead on Journal entries for internal and client-side meetings
Action Items
- Team to read up on GAN in medical fields
- Team to fully review previous COMP523 work – documentation, code, presentations, etc…